EP4695822A1 - Pre-treatment prediction of the response of cancer to neoadjuvant therapy - Google Patents

Abstract

Predicting a patient-specific therapeutic response prior to neoadjuvant treatment in a subject having cancer using a combination of machine learning operations and biology-based mathematical model. The combination provides higher accuracy of the predicted patient-specific therapeutic response using scans of the patient prior to any neoadjuvant treatment.

Description

PRE-TREATMENT PREDICTION OF THE RESPONSE OF CANCER TO NEOADJUVANT THERAPY

Cross Reference to Related Applications

[0001] This application claims priority to, and the benefit, of U.S. Provisional Patent Application Serial No. 63/495,874, filed April 13, 2023, which is incorporated by reference herein in its entirety.

Government License Rights

[0002] This invention was made with government support under grant nos, UO 1

CA142565, U01 CA174706, U24 CA226110 awarded by the National Institutes of Health and grant nos. CCF2019844 and DGE2137420 awarded by the National Science Foundation. The government has certain rights in the invention.

Background

[0003] In patients with cancer, neoadjuvant therapy is often used to treat tumors before surgery to reduce the tumor size and improve survival results. The predictive outcomes of neoadjuvant therapy can play an important role in determining which therapeutic regimen will provide the optimal course of action for treating cancer masses for the patient

[0004] Several frameworks have been developed to analyze the predictive outcomes of a neoadjuvant systematic therapy using radiographic scans to leant the features of the tumor. However, treatment decisions for various patient populations are based on population averages. Population-based approaches, which rely exclusively on statistical inference from properties of large populations as a whole, inevitably obscure conditions specific to the individual patient, especially for a disease as heterogeneous as cancer. Furthermore, limited established techniques are available for predicting the response of cancer before commencing neoadjuvant therapy, all of which require at least some treatment to occur before a response can be predicted.

[0005] Thus, there is a benefit to improving the predictive capability of cancer therapy outcomes before exposing the patient to any treatment.

Summary

[0006] An exemplary method arid system are disclosed for predicting a patient-specific therapeutic response prior to neoadjuvant treatment in a Subject having cancer using a combination of machine learning operations and a biology-based mathematical model The combination provides higher accuracy of the predicted pathological complete response using scans of th e patient prior to any neoadjuvant treatment. The exemplary method and system can thus better inform clinicians of better treatment options for the patient without having to expose the patient (or subject) to any specific treatments.

[0007] in an aspect, a method for predicting a patient-specific therapeutic response to neoadjuvant treatment in a subject having cancer is disclosed. The method includes: receiving imaging data (e.g., a plurality of images) associated with one or more imaging biomarkers of tumor cells in the subject, the imaging data being acquired prior to any neoadjuvant treatment therapy of a cancer; determining model parameters (e.g., tumor cell diffusion, net proliferation rate, absolute growth rate, and absolute death rate) of the tumor cells for a biology-based mathematical model by applying the imaging data to an artificial intelligence (Al) model; and determining one or more predictive tumor metrics (e.g., changes in tumor cellularity and/or tumor volume) using the biology-based mathematical model, wherein the biology-based mathematical model is configured to evaluate a spatiotemporal growth of the tumor cells in the subject using the model parameters of the tumor cells; wherein the one or more predictive tumor metrics are subsequently used to determine a predicted pathological complete response (pCR) to adjust or direct neoadjuvant therapy in the subject [0008] In some implementations, the one or more predictive tumor metrics are subsequently used to direct a course of neoadjuvant treatment in the subject

[0009] In some implementations, the one or more predictive tumor metrics are subsequently used to compare relative efficacies between two or more courses of neoadjuvant treatment, to subsequently direct neoadjuvant therapy in the subject.

[0010] In some aspects, the imaging data comprises data (e.g., one or more images) obtained from a magnetic resonance imaging (MRI) system, an ultrasound imaging system, a contrast enhanced ultrasound imaging system, a contrast enhanced mammography (CEM ) system, an X-ray computed tomography (CT) imaging system, a positron emission tomography (PET) imaging system, or combinations thereof.

[0011] I nIn some implementations, the imaging data comprises magnetic resonance imaging (MRI) data (e.g., a plurality of multiparametric MRI images).

[0012] In some implementations, the imaging data comprises diffusion-weighted imaging (DWI) data (e.g., a diffusion tensor model of diffusion-weighted imaging data) acquired from an MR! scanner. [0013] In some implementations, the imaging data comprises one or more dynamic contrast-enhanced (DCE) MRI (e.g., a plurality of sequential DCE-MRI frames).

[0014] In some implementations, the imaging data comprises an apparent diffusion coefficient (ADC) map.

[0015] In some implementations, the imaging data comprises one or more T₁-weighted images and/or T₁-weighted images.

[0016] In some implementations, the trained Al model determine tshe one or more predictive tumor metrics (e.g., changes in tumor cellularity and/or tumor volume) using the imaging data and clinical data (e.g., genomic data) from the subject. In some implementations, the genomic data includes gene expression levels (e.g., obtained from RNAseq) of a genetic biomarker corresponding to a clinically relevant effect on at least one model parameter of the tumor cells. In some implementations, the genetic data includes expression levels of one or more of PTEN, PIK3CA, AKT 1 , TP53, KRAS, MYC, and RB I . [0017] In some implementations, the clinical data (e.g., genomic data) is provided as a local input at each of a plurality of splits of the Al model. However, in other implementations, tire clinical data (e.g., genomic data) is provided as a global input to the Al model.

[0018] In some implementations, the method further includes spatially aligning two or more images of the imaging data (e.g., by determining a spatial offset or transformation of one image to another) at each of a plurality of voxels prior to determining the model parameters using the integration of the biology-based model with the Al model.

[0019] In some implementations, the Al model generates a spatially localized coefficient of the model parameters at each of the plurality of voxels.

[0020] In some implementatio tnhse, trained Al model comprises an encoder-decoder architecture (e.g., U-Net, ResNet, Multilayer Perceptron, SegNet, Fully Convolutional Networks (FCN), Mask R-CNN, Transformer, Diffusion model, Foundation model, Generative Adversarial Network (GAN), Long short-term memory network (LSTM ) or a combination thereof) that takes as input one or more of the imaging data and returns one or more of tumor proliferation rate and predictive tumor metrics (e.g., changes in tumor cellularity and/or tumor volume). In some examples, the Al model comprises a U-Net architecture.

[0021] In various implementation thse, biological computational model comprises a reaction-diffusion model. In some examples, the biology-based mathematical model is further configured to evaluate effectiveness of a course neoadjuvant treatment (e.g., by incorporating a treatment-induced death factor determining a death rate of the tumor cells based on a concentration, spatial distribution, and efficacy of a therapeutic agent).

[0622] In another aspect, a system is disclosed comprising one or more processors; and a memory having instructions stored thereon, wherein the instructions, when executed by the one or more processors, cause the one or more processors to perform any one of the above- discussed methods.

[0023] In various aspects, flic cancer is prostate cancer, breast cancer (e.g., triple-negative breast cancer (TNBC)), brain cancer, ovarian cancer, head and neck cancer, pancreatic cancer, cervical cancer, rectal cancer, esophagus cancer, liver cancer, stomach cancer, testicular cancer, vaginal cancer, uterine cancer, vulvar cancer, paranasal cancer, oropharyngeal cancer, or laryngeal cancer. For example, in some aspects, the cancer is locally advanced triple-negative breast cancer (TNBC).

[0024] In another aspect, a system is disclosed, having one or more processors; and a memory having instructions stored thereon, wherein the instructions, when executed by the one or more processors, cause the one or more processors to perform any one of the above- discussed methods.

[0625] In another aspect, a non-transitory computer-readable medium is disclosed, having instructions stored thereon, wherein the instructions, when executed by one or more processors, cause the one or more processors to perform any one of the above-discussed methods.

[0026] Additional advantages of the invention will be set forth in part in the description which follows, and in part will be clear from the description or may be learned by practice of the invention. The advantages of the invention will be realized and attained by means of the elements and combinations particularly pointed out in the appended claims. It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention, as claimed.

Brief Description of the Drawings

[0027] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments and, together with the description, serve to explain the principles of the methods and systems. The patent or application file contains at least one drawing executed in color. [0028] Figs. 1 A-1 B each depict a diagram of an example system configured with a neural network-integrated with a biology-based computational model that can predict the outcome of a course of neoadjuvant therapy before treatment has begun in accordance with an illustrative embodiment.

[0029] Fig. 2 shows an example method to predict the outcome of a neoadjuvant before treatment using imaging data in accordance with an illustrative embodiment.

[0030] Figs. 3A-3C show various implementations of the core CNN architecture used in a study to obtain model parameter predictions for a biology-based mathematical model using imaging data. Parameters c and x control the number of channels for each feature map in the architecture. The architecture includes a forward run of the biology-based mathematical model to also obtain predictions of cellularity at visits 2 (Fig. 3B) and 3 (Fig. 3C).

[0031] Figs. 4A-4C depict: (Fig. 4A), Cross-validation scheme used to train the network. (Fig. 4B), Network imaging inputs, core architecture, and outputs, including net proliferation rate prediction and V2/3 cellularity maps. (Fig. 4C), Loss calculation illustrating the use of normalized root mean square error (NRMSE) and concordance correlation coefficient (CCC) losses and the combination of loss on the parameter reconstruction and loss on the forward run cellularity predictions.

[0032] Figs. 5A-5B show (Fig. 5A) Architecture option one in which the genetic data input was incorporated after the expanding and contracting paths. (Fig. 5B) Architecture option two in which genetic data was included as constant inputs to the encoder.

[0033] Fig. 6 shows imaging data used in the experimental study as an input to the CNN architecture to provide the calibrated coefficients, which are later employed in the biologybased mathematical model.

[0034] Figs. 7A-7B show cellularity predictions through time for a patient who attained a pathological complete response (pCR) to therapy (Fig. 7A) and a patient who did not (non- pCR) (Fig. 7B).

[0035] Fig. 8 depicts validation plots comparing various predicted tumor metrics (changes in total tumor cellularity (TTC) and changes in total tumor volume (TTV)) for subsequent visits on both deep learning and calibration methodologies,

[0036] Fig. 9 depicts treatment, imaging, and pathology schema for the patient dataset in Example 2. (A) V1 V2, and V3 refer to the three imaging visits - pre-treatment (V1) post two cycles NAC (V2), post four cycles NAC (V3). (B) Overview of the modeling pipeline. The green boxes summarize the steps used to obtain and prepare the data, followed by calibration of the biology-based mathematical model (yel low) on all available imaging data. A convolutional neural network (CNN) was then trained using the calibrated parameters from the biology-based mathematical model (blue). Finally, the study evaluated the quality of the CNN predictions as compared to the biology-based model calibration and the measured data. Panels (C) and (D) provide more details on the steps involved and data required in calibrating the mathematical model and training the CNN, respectively. It is noted that both the biologybased mathematical model in (C) and the CNN in (D) predict the spatio-temporal response of a patient to neoadjuvant chemotherapy (NAC). These predictions can then be dtrecdy compared to the measured tumor cellularity maps at V2 and V3. It is also noted that the CNN can provide a prediction using the imaging data acquired before initiating treatment, whereas the biology-based model can provide a prediction using data from an individual patient [0037] Fig. 10 depicts an overview of convolutional neural network (CNN) architecture according to one implementation. (A) There are six channels of imaging input; one apparent diffusion coefficient (ADC) map, one pre-contrast Ti map, and four averages of 40 seconds of the dynamic contrast enhanced (DCE-) MRI percent enhancement (PE) timecourse. The seventh input channel is the location of the slice within the tumor. The input follows the concatenated encoder and decoder architectures followed by a linear layer to obtain the net proliferation rate prediction, k. Then, k is used in the biology-based mathematical model (BBM) to obtain the cellularity maps at V2 and V3 (N(x,1=V 2)d MAV=V3), respectively). In (B), it is shown how these three outputs are used in the loss function; specifically, k is used in weighted mean absolute error (wMAEk) and weighted concordance correlation coefficient (wCCG) loss calculations, as balanced by the hyperparameter y. The V2 and V3 cellularity maps are both used in the weighted concordance correlation coefficient loss Hyperparameter p balances loss between prediction of 1c versus and predictions of cellularity. [0638] Fig. 11 depicts representative test patients that did (left column) and did not (right column) attain a pCR after the completion of NAC. (A) and (B) show voxel-wise cellularity maps from the center slice of the tumor at V 1 through time to V2 and V3 for the measured data (top row), calibrated biology-based mathematical model (CBBM) using VI -V3 data (middle row), and the CNN prediction using V 1 data (bottom row). The CNN predicts the parameters that were calibrated to the biology-based mathematical model. (C) and (D) show three dimensional renderings of the measured, CBBM, and CNN predictions of tumor volume at V2 and V3. In particular, (C) is a case where the CNN fails to predict that the tumor will be gone at V3, while the CBBM correctly determines this, while (D) shows the CNN and CBBM both correctly determining that there will be residual tumor at. V3.

[0039] Fig, 12 depicts patient cohort results at V3 (end of A/C treatment). (A) shows the correlation in change in total tumor ceHularity (TTC) from VI to V3 between the calibrated biology-based model (CBBM) on VI-V3 data and the measured data (CCC - 0.98). (B) shows the same correlation, but between the CNN predictions from VI and the measured data. As a prediction was obtained for each test patient (« == 24) from each of the five folds of the cross validation, the median and range were plotted across the cross validation. The CCC is determined by comparing the measured change in TTC and the median values for tire CNN predicted changes in TTC (CCC “ 0.95), (C) shows the median and interquartile range (1QR) of tite absolute, voxelwise difference in tumor ceHularity (DTC) between the calibrated biology-based model and measured data. (D) shows the same metric as (C) for tire CNN prediction.

Detailed Description

[0040] Each and every feature described herein, and each and every combination of two or more of such features, is included within the scope of the present invention, provided that the features included in such a combination are not mutually inconsistent.

[0041] Definitions

[0042] As used herein, the term “cancer” includes any form of cancer, including but not limited to, solid tumor cancers such as prostate cancer, breast cancer (e.g., triple-negative breast cancer (TNBC)), brain cancer, ovarian cancer, head and neck cancer, pancreatic cancer, cervical cancer, rectal cancer, esophagus cancer, li ver cancer, stomach cancer, testicular cancer, vaginal cancer, uterine cancer, vulvar cancer, paranasal cancer, oropharyngeal cancer, or laryngeal cancer).

[0043] For example, the disclosed systems and methods can be used to predict neoadjuvant therapy outcomes for various types of breast cancar. The term “breast cancer” broadly encompasses any type of breast cancer that can develop in a subject. For example, the breast cancer may be characterized as Luminal A (ER+ and/or PR*, HER2-, low Ki67), Luminal B (ER* and/or PR*, HER2+(or HER2- with high Ki67), Triple negative, basal-like (ER—, PR—, HER2-) or HER2 type (ER™, PR—, HER2*).

[0044] As used herein, the term “neoadjuvant therapy" refers to therapy or treatment prior to surgery or other subsequent therapy. Examples of neoadjuvant therapy include chemotherapy, radiation therapy, hormone therapy, immunotherapy, and targeted therapy. [0045] As used herein, the term “subject” refers to any animat (e.g., a mammal), including, but not limited to, humans, non-humah primates, rodents, and the like.

Typically, the terms “subject” and “patient” are used interchangeably herein in reference to a human subject.

[0046] The term “imaging data” includes images and data acquired directly from an imaging apparatus (magnetic resonance imaging (MRI) system, an ultrasound imaging system, a contrast-enhanced ultrasound imaging system, a contrast-enhanced mammography (CEM) system, an X-ray computed tomography (CT) imaging system, a positron emission tomography (PET) imaging system) as well as any data images that are processed using mathematical and statistical methods known in the art and described herein. As used herein, the term “image" can refer to a two- or three-dimensional image.

[0047] As used herein, the term “biomarker” refers to an objective medical sign that is a measurable and quantifiable indicator of a physiologic or pathologic state of a tumor mass, defined structure, and/or living organism. A biomarker can be derived from any substance, structure or process that can be measured in the body or its products and influence or predict the incidence or outcome of disease. As an example, a biomarker may be used to identify individuals who are more likely to experience a favorable or unfavorable response to an intervention, medical product, or environmental exposure compared with individuals without the biomarker.

[0048] As used herein, the term “genomic data” includes genomic biomarkers and refers to a measurable DMA or RN A characteristic that is an indicator of normal biologic or pathogenic processes or a response to therapeutic or other interventions.

[0049] The term “imaging biomarker” refers to a biomarker that is present in, or derived from, imaging data and can be measured and quantified from tile imaging data to determine an indicator of a physiologic or pathologic state of a mass, defined structure and/or living organism within a region of interest, for which the imaging data is obtained.

[0050] The term “biology-based mathematical model” and the like refers to a system that uses mathematical concepts and language to describe one or more physiological processes of an organism. For example, a “biology-based mathematical model" can incorporate biological knowledge and experimentally determined parameters to simulate the growth of tumor cells. The model uses one or more mathematical equations to describe the behavior of cancer cells, taking into account factors such as cell proliferation, cell death, and interactions with the tumor microenvironment. The model can be calibrated to an individual patient via the inclusion of parameters such as cell proliferation rates and/or gene expression levels to improve its accuracy and predictive power.

[0051] All patents and publications mentioned in the specification are indicative of die levels of skill of those skilled in the art to which the invention pertains. References cited herein are incorporated by reference herein in their entirety to indicate the state of the art as of their filing date, ami it is intended that this information can be employed herein, if needed, to exclude specific embodiments that are in the prior art.

[0052] Example System

[0053] Figs. 1A and IB each show a diagram of an example system 100 (shown as 100a, 100b, respectively) configured with a neural network that can accurately and reliably predict the outcome of neoadjuvant before treatment begins in accordance with an illustrative embodiment

[0054] In the example shown in Fig. 1 A, the system 100a includes a framing network

110a and a production network 120a. The training network 110a includes a neural network 114 that receives imaging data 112a from a data store 101. In various implementations, the neural network comprises an encoder-decoder architecture (e.g., U-Net, ResNet, Multilayer Perceptron, SegNet, Fully Convolutional Networks (FCN), Mask R-CNN, Transformer, Diffusion model. Foundation model, Generative Adversarial Network (GAN), Long shortterm memory network (LSTM) or a combination therein) configured to return one or more of a tumor proliferation rate and predictive tumor metrics (e.g., changes in tumor cellularity and/or tumor volume). In some implementations, the neural network is a U-Net architecture. [0055] The imaging data 112a (e.g., a plurality of images) is associated with one or more imaging biomarkers of tumor cells from a population of patients having cancer. In various implementations, the imaging data includes data, such as one or more images, obtained from a magnetic resonance imaging (MRI) system, an ultrasound imaging system, a contrast- enhanced ultrasound imaging system, a contrast-enhanced mammography (CEM) system, an X-ray computed tomography (CT) imaging system, a positron emission tomography (PET) imaging system, or combinations thereof. In some examples, the imaging data includes magnetic resonance imaging data (e,g., a plurality of multiparametric MRI images). In various implementations, the imaging data comprises diffusion-weighted imaging (DWI) data (e.g., a diffusion tensor model of diffusion-weighted imaging data) acquired from an MRI scanner. In some examples, the imaging data comprises one or more dynamic contrast- enhanced (DCE) MRI (e.g., a plurality of sequential DCE-MRI frames). In some implementations, the imaging data comprises an apparent diffusion coefficient (ADC) map. In some implementations, the imaging data comprises one or more //-weighted images and/or T>-weighted images.

[0056] The neural network 114 is trained using the imaging data 112a to determine model parameters 116 (e.g., tumor cell diffusion, net proliferation rate, absolute growth rate, and absolute death rate) of tumor cells in a population of patients from the imaging data. The system 100a further includes a biology-based mathematical model 118, which is calibrated using the model parameters 116. The training operation after the fully connected layer 115, in the provided example, is configured to employ conventional operations, e.g., according to U- Net, e.g., employing gradient descent and various normalization operations, and thus are not further described herein. Other training operations may be employed. Figs. 4A-4C provides an example of a cross-validation scheme to train a CNN according to the present disclosure. [0057] Referring now back to Fig. I A, the production network 120a includes a trained neural network 124 configured to receive imaging data 122a. The imaging data 122a is retrieved from a data store 111.

[0058] The imaging data 122a (e.g., a plurality of images) is associated with one or more imaging biomarkers of tumor cells in a subject having cancer or suspected of having cancer. In various implementations, the imaging data includes data, such as one or more images, obtained from a magnetic resonance imaging (MRI) system, an ultrasound imaging system, a contrast-enhanced ultrasound imaging system, a contrast-enhanced mammography (CEM) system, an X-ray computed tomography (CT) imaging system, a positron emission tomography (PET) imaging system, or combinations thereof. In some examples, the imaging data includes magnetic resonance imaging data (e.g., a plurality of multiparametric MRI images). In various implementations, the imaging data comprises diffusion-weighted imaging (DW1) data (e.g., a diffusion tensor model of diffusion-weighted imaging data) acquired from an MRI scanner. In some examples, the imaging data comprises one or more dynamic contrast-enhanced (DCE) MRI (e.g., a plurality of sequential DCE-MRI frames). In some implementations, fee imaging data comprises an apparent diffusion coefficient (ADC) map. In some implementations, fee imaging data comprises one or more /'/-weighted images and/or T₂-weighted images.

[0059] When fee trained neural network 124 receives the imaging data 122a, the trained neural network outputs one or more model parameters 126 (e.g., tumor cell diffusion, net proliferation rate, absolute growth rate, and absolute death rate) of fee tumor cells. The model parameters 126 determined by the trained neural network 124 are used to calibrate a biologybased mathematical model 128. Although the biology-based mathematical model 128 shown in Fig. I A includes a reaction-diffusion model, other models can also be used. The biologybased mathematical model 128 is configured to evaluate a spatiotemporal growth of die tumor cells in the subject using the model parameters 126 of the tumor cells and determine one or more predictive tumor metrics 130. In some implementations, the biology-based mathematical model can further be configured to evaluate an effectiveness of a course neoadjuvant treatment (e.g., by incorporating a treatment-induced death factor determining a death rate of the tumor cells based on a concentration, spatial distribution, and efficacy of a therapeutic agent).

[0060] The one or more predictive tumor metrics 130 are subsequently used to determine a predicted pathological complete response (pCR) 132 to adjust or direct neoadjuvant therapy in the subject. In various implementations, the one or more predictive tumor metrics are subsequently used to direct a course of neoadjuvant treatment in the subject In some implementations, the one or more predictive tumor metrics are subsequently used to compare relative efficacies between two or more courses of neoadjuvant treatinent to subsequently direct neoadjuvant therapy in the subject

[0061] Fig. 1 B shows another diagram of an example system 100b configured with a neural network that can accurately and reliably predict the outcome of neoadjuvant before treatment begins in accordance with an illustrative embodiment

[0062] In the example shown in Fig. IB, the system 100b includes a training network 110b and a production network 120b. The training network 110b includes a neural network 114 that receives imaging data 112a and clinical data 112b (e.g., genomic data) from a data store 101.

[0063] The neural network 114 is trained using the imaging data 112a and clinical data 112b (e.g., genomic data) to determine model parameters 116 (e.g., tumor cell diffusion, net. proliferation rate, absolute growth rate, and absolute death rate) of tumor cells in a population of patients from the imaging data 112a. The system 100b further includes a biology-based mathematical model 118, which is calibrated using the model parameters 116. The training operation after the fully connected layer 115, in the provided example, is configured to employ conventional operations, e.g., according to U-Net, e.g., employing gradient descent and various normalization operations, and thus are not further described herein. Other training operations may also be employed. [0064] The production network 120b includes a trained neural network 124 configured to receive imaging data 122a and clinical data (e.g., genomic data) 122b. The imaging data 122a and clinical data 122b are retrieved from a data store 1 1 L

[0065] The imaging data 122a (e g., a plurality of images) is associated with one or more imaging biomarkers of tumor cells in a subject having career or suspected of having cancer. The clinical data 122b includes patient specific information concerning or influencing the health status of the patient. The clinical data can include, but is not limited to, age, sex, weight, genomic data, etiopathology data, anamnesis data, menopausal/hormonal status, data obtained by in vitro diagnostic methods such as blood or urine tests, or a combination thereof. In various aspects, tire clinical data includes genomic data. The genomic data can include, for example, gene expression levels (eg., obtained from RNAseq) of a genetic biomarker corresponding to a clinically relevant effect on at least one model parameter of the tumor cells. In some examples, the clinical data includes histoimmunological data.

[0066] When the trained neural network 124 receives tire imaging data 122a and clinical data 122b, the trained neural network outputs model parameters 126 (e.g., tumor cell diffusion, net proliferation rate, absolute growth rate, and absolute death rate) of the tumor cells. The model parameters 126 determined by the trained neural network 124 are used to calibrate a biology-based mathematical model 128. The biology-based mathematical model 128 is configured to evaluate a spatiotemporal growth of the tumor cells in the subject using the model parameters 126 of the tumor cells and determine one or more predictive tumor metrics 130. The one or more predictive tumor metrics 130 are subsequently used to determine a predicted pathological complete response (pCR) 132 to adjust or direct neoadjuvant therapy in the subject

[0067] The exemplary system and method can be based on any type of machine learning or artificial intelligence system, including, and not limited to, convolutional neural networks, autoencoders, recombinant neural networks, and long-term short-term memory (LSTM) networks. The system (e.g., 100a, 100b) can be implemented via a processing unit (e.g., processor) configured to execute computer-readable instructions. In its most basic configuration, the processing unit includes at least one processing circuit (e.g., core) and system memory. Depending on the exact configuration and type of computing device^ system memory may be volatile (such as random-access memory (RAM)), non-volatile (such as read-only memory (ROM), flash memory, etc.), or some combination of the two. The processing unit may be a standard programmable processor that performs arithmetic and logic operations necessary for the operation of the computing device. As used herein, processing unit and processor refers to a physical hardware device that executes encoded instructions for performing functions on inputs and creating outputs, including, for example, but not limited to, microprocessors (MCUs), microcontrollers, graphical processing units (GPUs), and application-specific circuits (ASICs).

[0068] While instructions may be discussed as executed by a processor, the instructions may be executed simultaneously, serially, or otherwise executed by one or multiple processors.The processing unit may also include a bus or other communication mechanism for communicating information among various components of the computing device. In some embodiments, the processing unit is configured with co-processors (e.g., FPGA, ASIC) or Al-processors.

[0069] Example Method

[0070] Fig. 2 shows a method 200 for predicting a patient-specific therapeutic response to neoadjuvant treatment in a subject having cancer is provided. The method 200 includes receiving (202) imaging data (e.g., a plurality of images) associated with one or more imaging biomarkers of tumor cells in the subject, the imaging data being acquired prior to any neoadjuvant treatment therapy of a cancer. In various implementations, the imaging data comprises data, such as one or more images, obtained from a magnetic resonance imaging (MRI) system, an ultrasound imaging system, a contrast-enhanced ultrasound imaging system, a contrast-enhanced mammography (GEM) system, an X-ray computed tomography (CT) imaging system, a positron emission tomography (PET) imaging system, or combinations thereof. In some examples, the imaging data includes magnetic resonance imaging data (e.g., a plurality of multiparametric MRI images). In various implementations, the imaging data comprises diffusion-weighted imaging (DWI) data (e.g., a diffusion tensor model of diffusion-weighted imaging data) acquired from an MRI scanner. In some examples, the imaging data comprises one or more dynamic contrast enhanced (DCE) MRI (e.g., a plurality of sequential DCE-MRI frames). In some implementations, the imaging date comprises an apparent diffusion coefficient (ADC) map. In some implementations, the imaging data comprises one or more '//-weighted images and/or //-weighted images. [0071] The method 200 further includes determining (204) model parameters of the tumor cells from the imaging data using a trained artificial intelligence (Al) model In various implementations, the model parameters include one or more tumor cell diffusion, net proliferation rate, absolute growth rate, and/or absolute death rate. The Al model of the present method 200 utilizes an image segmentation model; however, other types of Al models, such as an image (egression model or an image classification model, can also be used. In some implementations, the trained Al model comprises an encoder-decoder architecture (e.g., U-Net, ResNet, Multilayer Perceptron, SegNet, Fully Convolutional Networks (FCN), Mask R-CNN, Transformer, Diffusion model, Foundation model, Generative Adversarial Network (GAN), Long short-term memory network (LSTM) or a combination thereof) that takes as input one or more of the imaging data and returns one or more of tumor proliferation rate and predictive tumor metrics (e.g., changes in tumor cellularity and/or tumor volume). In some examples, the Al model comprises a U-Net architecture.

[0072] Method 200 further includes determining (206) one or more predictive tumor metrics using a biology-based mathematical model. The biology-based mathematical model is configured to evaluate the spatiotemporal growth of the tumor cells in the subject using the model parameters of the tumor cells. In various implementations, the one or more predictive tumor metrics include changes in tumor cellularity and/or tumor volume.

[0073] In some examples, the trained Al model determines the one or more predictive tumor metrics (e.g., changes in tumor cellularity and/or tumor volume) using the imaging data and clinical data (e.g., genomic data) from the subject. For example, in some implementations, the genomic data includes gene expression levels (e.g„ obtained from RNAseq) of a genetic biomarker corresponding to a clinically relevant effect on at least one model parameter of the tumor cells. In various examples, the clinical data (e.g., genomic data) is provided as a local input at each of a plurality of splits of the Al model. However, in other implementations, the clinical data (e.g., genomic data) is provided as a global input to the Al model. In some implementations, the method further includes spatially aligning two or more images of the imaging data (e.g., by determining a spatial offset or transformation of one image to another) at each of a plurality of voxels prior to determining the model parameters using the integration of the biology-based model with the Al model. In some examples, the integrated biology-based and Al model generates a spatially localized coefficient of the model parameters at each of the plurality of voxels.

[0074] In various implementations, the biological computational model comprises a reaction-diffusion model.

[0075] Method 200 also includes adjusting or directing (208) neoadjuvant therapy in the subject based on the predicted pathologically complete response. In various implementations, the one or more predictive tumor metrics are subsequently used to direct a course of neoadjuvant treatment in fee subject. In some implementations, fee one or more predictive tumor metrics are subsequently used to compare relative efficacies between two or more courses of neoadjuvant treatment, to subsequently direct neoadjuvant therapy in the subject

[0076] Examples of adjustments may include adjustments to (i) chemotherapy or (if) therapy that targets Human Epidermal Growth Factor Receptor 2 (HER2), Estrogen Receptor (ER), or Progesterone Receptor (PgR).

Experimental Resalts and Examples

[0077] Example #1: MRI-based digital models forecast patient-specific treatment responses to neoadjuvant chemotherapy in triple-negative breast cancer.

[0078] A study was conducted to determine a patient-specific therapeutic response to neoadjuvant treatment using machine learning operations in combination wife a biologybased mathematical model to determine a predicted pathological complete response for a subject or patient solely using scans of the patient prior to any neoadjuvant treatment .

[0079] Equation (I) shows a biology-based, reaction-diffusion-based model used in the study to evaluate predictive tumor metrics of a patient [0080] In Equation 1, Z)(x,/) is the mechanically coupled diffusion coefficient, A(x) is the proliferation rate at each voxel, theta is the tumor cell carrying capacity per voxel, and A_n(x, t) is a function describing the drug-induced death of tumor cells.

[008] ] lire first term on fee right-hand side of fee equation describes fee diffusion, or movement, of tumor cells. If D(x, t) is set to a constant value, it is assumed feat diffusion is free, or not subject to mechanical effects. Alternatively, this term can be mechanically coupled to allow fee tumor cell diffusion to account for fee compression of surrounding tissues using Equation (2).

[0082] In Equation 2, D_o is the diffusivity coefficient in the absence of external forces, as would be used, if setting this term to a constant value. In the mechanically coupled case, however, this value was dampened with the exponential term to reduce the cell mobility based on the stiffness of surrounding tissue quantified by the von Mises stress (x, t)) and an empirical coupling constent y. To obtain von Mises stress, the mechanical equilibrium was defined as Equation (3). [0083] In Equation 3, a is the Cauchy stress tensor and Xf is a tumor cell-force coupling constant. To implement this, Equation (3) was rewritten in terms of displacement u under an assumption of a linear, elastic, isotropic material, as seen in Equation (4): [0084] In Equation 4, v is Poisson’s ratio and G is the shear modulus defined as G = Ef (2(1 — v)) with Young’s modulus E. Young’s modulus varies spatially between tumor, fibroglandular, and adipose tissue.

[008$] From the Cauchy stress tensor <r, the von Mises stress was calculated using Equation. [0086] The mechanical coupling was implemented using methods analogous to those found in Hormuth et al (2018).

[0087] The second term on the right-hand side of this equation describe tshe proliferation of tumor cells via logistic growth with local proliferation rate k(x) and carrying capacity fl. A local proliferation rate means that a proliferation value was determined for each voxel in a patient’s tumor. A global proliferation rate, however, refers to a single value assigned to all voxels in a tumor. The third term on the right-hand side of this equation describes the treatment-induced death of tumor cells through t), which was determined using Equation (6).

[0088] In Equation 6, a_n was defined as a drag efficacy parameter Adriamycin, Cytoxan, or Taxol with n = 1 ,2,3, respectively . The study obtained drug concentration caused by the therapy administration from the normalized area under the DCE-MRI enhancement curve. The study dampened this by the exponential decay of each drug in the body described by decay rates for Adriamycin and Cytoxan, and Taxol a, respectively, with ranges taken from clinical trial literature. Barpe et al. (2010); Zhu et al. (2015); Powis etal. (1987); Mori et al. (2006).

[0089] Using Equations (1 ) - (6) in combination with the patient imaging data, the study calibrated a subset of the parameters, such as on the computational domain of the segmented breast contour, The study used the cellularity timecourse obtained from sequential DW-MRI data sets obtained at the pre-treatment visit VI and at least one of the on-treatment visits V2 (and/or V3) and the Levenberg-Marquardt algorithm to obtain these values. The experiment provided further details on the values and calibration ranges for all model parameters in Table 1. Figs. 7A-7B show cellularity predictions through time for a patient who attained a pathological complete response (pCR) to neoadjuvant therapy (Fig.

7A) and a patient who did not (non-pCR) (Fig. 7B). Fig. 8 depicts validation plots comparing various predicted tumor metrics (changes in total tumor cellularity (TTC) and changes in total tumor volume (TTV)) for subsequent visits on both deep learning and calibration methodologies.

Table 1: Model parameters and units. [0090] Convolutional Neural Network to Generate the Model Parameters of the Tumor

[0091] The study employed a U-Net-inspired convolutional neural network (CNN) architecture designed to relate pretreatment data to calibrated model parameters in a version of the mechanism-based model described above. The core network architecture is shown in Figure 3A-3C and described below.

[0092] Inputs to the U-Net-inspired convolutional neural network (CNN! The study used the following inputs derived from MRI data:

[0093] ADC map calculated from DW-MRk The ADC map was selected as this parameter has been shown to be inversely proportional to celhdarity. The study normalized tumor ADC to a range of [0,1 ] using the ADC of water at 3TC (3x1 O'⁵ mmV¹).

[0094] Averages of DCE percent enhancement timecourse frames from DCE-MRI; The study took timepoint averages to increase the signal-to-noise ratio (SNR). The average of the first five post-contrast frames of the time course was taken to provide information on washin. The five frames from 60 to 100 seconds and the five frames from 150 to 190 seconds were averaged as the 60 and 150 second frames have previously been shown to provide accurate tumor volume estimates. The study averaged the last five frames to provide information on wash-out These inputs were then normalized to the maximum PE described later herein.

[0095] Not all of the frames for the entire time course were used as individual CNN inputs. In the standard-of-care setting, only three DCE-MRI time points (pre-contrast, early post-contrast, and late post-contrast) were typically measured. Thus, the use of fewer input times can be advantageous for future widespread applications. Additionally, having a very large quantity of input images to the CNN can lead to difficulty in generalizability. Without enough data to realize a relationship between all input images and the output, the network will instead memorize the training data. With the amount of data available in the study, it was difficult to construct a network that could generalize between an input, including all PE frames from the DCE-MRI data, and the output

[0096] Ti map from variable flip angle Ti mapping: Ti relaxation time varies between different types of tissue, so the study selected the Ti map to provide information on tissue heterogeneity across the tumor. These Ti maps were normalized to the maximum Ti relaxation time. Fig. 6 shows example images depicting several types of imaging data used in the study. [0097]The presently described CNN operated in two dimensions on slices of patient tumors, so the study also included input of slice location along the third dimension. For each slice, the minimum distance from the slice to either the top or bottom of the tumor was determined. The slice location was determined as the natural logarithm of this distance. The study normalized this input for each patient by dividing by the maximum slice location value in the patient’s tumor. All imaging inputs were cropped to 32^32 voxels encompassing each patient’s tumor.

[0098] Outputs of the convolutional neural network (CNN). The CNN determined a relationship between inputs (i.e., ADC', 7/ maps, DCE PE) and calibrated model parameters in a model based on Equation (I) (i.e.. Do, k, «) specific to each aim. Thus, the output of the CNN was a prediction of the model parameters (i.e., Do, k, a) that was calibrated to Equation (1 ). As exhibited in Figures 3 A-3C, the calibrated model parameter prediction was used in a forward run of a mathematical model (i.e., Equation (1)). From this, the study also obtained predictions of cellularity at V2 and V3. Using the cellularity predictions when training the CNN, the study can calculate the error not only on the network’s ability to correctly predict model parameters but also on the quality of a forward run of Equation (1) using those determined model parameters. This means that the CNN was not restricted to just reconstructing the calibrated parameters, it can also adjust them based on the error in the predicted cellularity from the forward run. The study allowed the network this flexibility in parameter predictions to aid in generalizability. That is, while the network predicts parameters from the calibration, the calibrated parameters did not perfectly reconstruct cellularity. The network can adjust the calibrated parameter values to those that provide a similar quality cellularity reconstruction.

[0099] Architecture. The study used the core CNN architecture seen in Figures 3A-3C, in which imaging inputs follow contracting and expanding paths characteristic of a U-Net with alterations specific to this work. The CNN architecttire was implemented in Python (3.10.6; Python Software Foundation, Wilmington, DE, USA) using the PyTorch package. Paszke et al. (2017).

[0100] The contracting path was composed of multiple sets of two padded 3x3 convolution layers with batch normalization (i.e., a technique that standardizes inputs to a layer for each batch) and the parametric rectified linear unit (PReLU) (He et al. (2015)) activation function (i.e., a function applied elementwise to a layer to determine its final output). Batch normalization increased the stability and speed of training through recentering and rescaling inputs. The study applied the batch normalization. Activation functions introduce nonlinearity to the network. The PReLU activation function was defined as Equation (7)i [0101] In Equation 7, a is a learned parameter. Padding the convolutions ensures the original input size was retained. The two convolutional layers increased the number of channels by a factor of two. The study further used a maxpool layer with a 2*2 kernel with a stride of two to decrease the height and width of the output of the double convolution by a factor of two while preserving the number of channels. The maxpool slid the 2x2 kernel across the image, moving two voxels with each slide, and formed a new image by taking the maximum of each region it slid over. The architecture used three of these double convolution structures, bringing the image down to a feature map of size 4x4 with eight times the number of channels in the first convolution layer. The contracting path can be thought of as an encoder architecture, as the convohition-maxpool blocks encode representative features of the input image.

[0102] The expanding path takes input of the feature maps learned by the expanding path.

It includes a bilinear upsample by a factor of two with PReLU activation, a concatenation with the correspondingly sized feature map from the contracting path, and two padded 3x3 convolutions that halve the number of channels in the image. Each convolution was followed by batch normalization and PReLU activation.The expanding path contains the same number of double convolution layers as the contracting path such that 32*32 voxel maps are outputted- the same size as the input The expanding path can be thought of as a decoder architecture, as it takes an input of the features obtained by the encoder and decodes them into a useful output. The output of the expanding path enters an architecture specific to each aim to obtain a prediction of model parameters to use in a forward run.

[0103] Results and Discussion. One purpose of the present experiment was to develop a methodology that allows for an accurate prediction of the eventual response of locally- advanced TNBC to NAT before initiating therapy. The study addressed this goal through a combination of convolutional neural networks (CNN) and mechanistic modeling. [0104] Biology-based modeling framework. To accomplish this, the study used a simplified version of the mechanistic model, as described in Equation (8). [0105] In Equation 8, Z> is a fixed tumor cell diffusivity (1 e-3 mnr/day), k is a globally calibrated net proliferation rate (i.e., implicitly capturing all proliferative and drug effects) in the range [*0.2, 0.1] 1/day, and 0 is a patient-specific carrying capacity (1.5 times the maximum cellularity at VI). The study chose to fix the diffusivity based on the result of a sensitivity analysis. While fee data was three-dimensional, this model was calibrated using the Levenberg-Marquardt algorithm and imaging timepoints VI -V3 in two dimensions on a slice-by-slice basis, meaning one net proliferation rate was obtained for each slice. The study calibrated slice-by-slice for several reasons: I) it provides more data for the neural network, 2) it eliminates the need to develop a method to compensate for different tumors having different numbers of slices, 3) it ensures that a single poor prediction from the neural network will have less impact on the overall result for a patient, and 4) it enables the capture of a degree of heterogene ity across fee tumor because k is fixed by slice rather than fixed over the whole tumor.

[0106] Discussion

[01071

[0108] Inputs and outputs. The neural network relates an input of pre-treatment imaging data to an output of calibrated net proliferation rate in fee biology-based model (Equation (8)). The study used the pretreatment imaging data and slice location described above as inputs. The CNN related this input to a single net proliferation value for each input slice. The net proliferation value was normalized to the range [0,1 ] using fee allowable parameter range from fee calibration via min-max normalization, as seen in Equation (9).

[0109] In Equation 9, k is the original value, k_WTm is the normalized value, and and kmis are the maximum and minimum allowable proliferation values in the calibration, respectively. Once the network predicts this parameter value, it uses it to forward run the mathematical model in Eq, (8) to output cellularity predictions for V2 and V3, giving the neural network three outputs (i.e., k, cellularity at V2, and cellularity at V3) to consider when training.

[0110] CNN architecture. As visualized in Figure 4B, the input to the convolutional neural network (CNN) was 32*32 voxels with seven channels - one for ADC, one for Ti map, four for DCE PE time course averages, and one for slice location. The input then follows the CNN architecture composed of a contracting and expanding path described above. To predict a global net proliferation, the final 32*32 output of this architecture was flattened and used as input to the linear layer. The linear layer outputted a single value with a sigmoid activation to obtain a final prediction of net proliferation in the range of [0,1 ]. The study then rescaled this prediction to the net proliferation range used for calibration by solving the min-max normalization Equation (9) for k and using the CNN prediction for k_norm. The net proliferation prediction was then used in a forward run of Equation (8) to provide predictions of cellularity for comparison to the measured cellularity at V2 and V3.

[0111] Training. Holding a randomly selected n - 14 patients as a test set, the study performed training on n = 100 patients in a random five-fold cross-validation, as seen in Figure 4A. This cross-validation was repeated eight times such that 1.1.2 of the 114 patients were included in one test set From the cross-validation, this means five net proliferation predictions were obtained for each of the 112 patients. The study employed a combination of grid searches over parameter pairs and toning of individual parameters to select hyperparameters. Test results were evaluated at the last epoch in which the validation loss decreased. Network weights were updated throughout training via a loss function that considers multiple metrics across the net proliferation reconstruction and cellularity map predictions, as seen in Figure 4C.

[0112] First, the loss was calculated on the net proliferation predictions. Here, the study used a combination of two metrics on the predicted versus calibrated net proliferation: 1) one minus concordance correlation coefficient (CCC), and 2) normalized root mean square error (NRMSE). The mean of the true values for normalization was used in the NRMSE. While the NRMSE loss ensures that the parameter predictions are close to the true values, the CCC loss ensures that the predictions follow tlie overall distribution of the calibrated net proliferation rates. The distribution of calibrated net proliferation values was not normal, having a sharp peak at the lower end of the distribution. By incorporating the CCC into the loss, the study showed an improved ability to capture this peak Second, the loss was calculated on the cellularity maps that were obtained from running the model in Equation (8) forward in time to V2 and V3. The study calculated CCC at V2 and V3 over the region of tumor-bearing voxels at VI separately, then averaged the result. The parameter reconstruction and cellularity prediction loss calculations were combined as follows in Equation ( 10).

[0113] In Equation 10 and that these coefficients were treated as network hyperparameters. [0114] Pretreatment predictive capabilities, 'Fhe study also sought to improve the pretreatment predictive capabilities of the combined mechanistic modeling and convolutional neural network framework by incorporating genetic data and using a more comprehensive mechanistic model.

[0115] Biology-based modeling framework. Using fee mechanistic model ftom above

(Equation 2) as a starting point, the experiment sought to characterize tumor response to therapy more comprehensively. This was done by splitting the net proliferation term into two terms - one describing cell proliferation and one describing cell death, as seen in Equation (11).

[0116] In Equation 11 , the equation has a fixed diffusivity D (le-3 mm²day⁴) and patient-specific carrying capacity <9(1.5* the largest cellularity at VI ). The two parameters were calibrated: 1) proliferation rate k in [0,0.1] 1/day, describing cell proliferation, and 2) drug efficacy a in [0,0.8) l/(pM day) describing the death of cells due to the therapy. The drug term uses a combined efficacy alpha for both drugs. The drug decay values, were fixed to the center of allowable ranges found in literature. Barpe et al. (2010); Zhu el al. (2015); Powis et al. (1987); Mori et al. (2006). To better caphire local changes in tumor cellularity, one of these parameters was calibrated locally, meaning a map of values was obtained rather than a single value for the whole tumor. Following from previous work proliferation rate A could be selected as fee locally calibrated parameter. However, a sensitivity analysis over D, k,a nd a, identified drug efficacy a as a significant parameter when it comes to quantifying changes in total tumor cellularity and volume. As such, the exploration of two calibrations was considered. 1) local k, global a, and 2) global k, local a. The locally calibrated parameter is referred to as pi(x\ and fee globally calibrated parameter asp?.

[0117] Machine learning framework [0118] Inputs and outputs. The presently described neural network sought to determine a relationship between inputs of pretreatment imaging and genetic data and outputs of calibrated proliferation rate and drug efficacy parameters. The study continued to use the pretreatment imaging and slice location data as the imaging input. For the genetic input, die study first identified a set of eight genes with known clinical relevance to TNBC. From this set, genes were selected with established effects on cell proliferation and death due to drugs - the two calibrated parameters in the model. The study selected PTEN, PIK3CA, AKT 1 , TP53, KRAS, MYC, and RBI as relevant to cell proliferation and PTEN, PIK3CA, AKTI, TP53, KRAS, MYC, and CCNE1 as relevant to cell death. The study obtained one gene expression level from the RNAseq data for each patient, thus the genetic input for all slices of a single patient was constant As outlined in Figures 5A-5B, two options were proposed for incorporating both global and local inputs in the architecture.

[6119] After the expanding and contracting paths, the study was split into two separate architectures - one to predict the local parameter />/(*)> and one to predict the global parameter p>, as shown in Figure 5A. A vector of genetic data was included as an input to each split of the architecture. This allows the set of identified genes for each parameter to only impact the prediction of that parameter. However, this means that the genetic data were not considered in most of the network architecture. Without wishing to be bound by theory, the addition of genetic data suggests an improvement to the predictive capabilities of the CNN; however, it may be limited to consider the genetic data in only a small portion of the architecture.

[0001] Similar to the incorporation of slice location, constant channels could also be included in the encoder input, as shown in Figure 5B. This is advantageous in its simplicity and in that the genetic data was considered through the entirety of the network. However, the study selected separate gene subsets for proliferation versus death due to drug, and this mefood would use all genes when making predictions for each parameter.

[0002] CNN Architecture. The network output now includes predictions of a locally calibrated parameter, a globally calibrated parameter, and the result of a forward run of Equation ( 11 ) to V2 and V3 using those parameters. The study established options for incorporating the global inputs above. The globally calibrated parameter was obtained using the same fully connected architecture developed above (Figures 4A-4B). Three options were additionally proposed for how the network could reconstruct the locally calibrated parameter: [0120] One option would be to have the network directly reconstruct the 32x32 parameter map. However, through investigating the locally calibrated parameter maps from previous work (Wu et al. (2022)) these maps were found to be heterogenous and often have large differences in parameter values between voxels. CNNs are better suited to smooth outputs, so this type of heterogeneity can be difficult for a CNN to capture.

[0121] The study could also use a principal component analysis (PCA) to reduce from 32x32 parameter maps to just a few principal components representing the map. The PCA would linearly transform the data to a new coordinate system in which a vast majority of the variation in the data could be captured in a smaller number of dimensions.

[0122] As the contracting path is designed to output a 32x32x0 feature map, the study can simply apply additional size-preserving convolution layers to bring this to a 32x32 prediction of the locally calibrated parameter pi(x). The same architecture found in Figure 4B could be used to obtain a prediction of the globally calibrated parameter pi. Finally, the study used both global and local parameter predictions in a forward run of the model (Equation (5)) to obtain the V2/3 cellularity predictions.

[0123] Examole #2; Inteeratine biology-based and data-driven modeline to predict the response of locally advanced triple-negative breast cancer before initiating neoadjuvant chemotherapy

[0124] In another aspect of the study, an integrated deep learning ami biology-based modeling was developed to obtain spatio-temporaliy resolved predictions of TNBC response to NAC before initiating therapy (Figure 9, panel B). To do so, a CNN was trained to predict interpretable, CBBM parameters representing tumor growth-response behaviors from the pretreatment MR images (Figure 9, panels C & D). The predicted parameters were used to estimate patient-specific treatment response in a test cohort It was believed that the concordance correlation coefficient between tile CNN predicted and measured change in total tumor cellularity (TFC) and total tumor volume (TTV) would be sufficiently high (i.e., > 0.9) and that a receiver operating characteristic (ROC) curve analysis using TTC and TTV to predict pCR status will attain an area under the ROC curve greater than random chance (ie., > 0.5). Through this process, the study combined deep learning and biology-based modeling to predict the response of locally advanced, TNBC before initiating NAC.

[0125] Materials and Methods

[0126] Patient data. All patients (n ::: 146) were women, and the median (range) age was 49 (28-78) years. Patients received multiparametric MRI scans before (VI ), after two cycles (V2), and after four cycles (V3) of neoadjuvant Adriamycin/Cytoxan (Figure 9, panel A). (See Wu et al. (2022) fbr example details regarding MRI acquisition.) Patients with >70% reduction in tumor volume at V3 (by ultrasound) then received 12 cycles of Taxol therapy, followed by surgery to determine pCR status.

[0127] Imaging data. The images collected at each visit included diffusion weighted MRI series with two //-values (100, 800 s/mm²), Ti-weighted series with four flip angles (3°, 5°, 10°, 15°), and a //-weighted dynamic contrast enhanced series with a median (range) of 42 (36-50) frames and a temporal resolution of I LI (7.6-12.4) seconds. The diffusion weighted MRI data yielded the apparent diffusion coefficient maps which were then converted to a cellularity map. Yankeelov et al. (2012). A T$ map was calculated from the pre-contrast variable flip angle data. The study interpolated all dynamic contrast enhanced MRI timecourses to 360 seconds with 10 seconds between each frame and calculated percent enhancement for each frame. Both inter- and intra-visit registration were performed to align all images to a common frame. Jarrett et at. (2021).

[0128] Mathematical model. The experiment used a simplified version of the biologybased mathematical model presented shown in Eq. (8), It describes the change in number of tumor cells, (mm) and time i (days): where is the fixed tumor celt migration rate, k (day*¹) is the calibrated net effect of cell proliferation and drug-induced dea th, and 0 (cells) is the fixed carrying capacity. For each patient, k was calibrated in Eq. (8) to the VI, V2. and V3 images on a slice-by-slice basis to obtain one value for each slice to capture tumor heterogeneity. Patients were excluded with lower quality calibrations from the CNN training. This method in which the study calibrated multiple timepointe of imaging data to Eq. (8) is referred to as the calibrated biology-based model (CBBM).

[0129] CNN inputs and outputs. The CNN relates pre-treatment imaging data to the

CBBM net proliferation rate, k, determined by calibrating pre- and on-treatment imaging data to Eq. (8). The CNN inputs were the apparent diffusion coefficient map, Ti map, and four time-averaged frames derived from the dynamic contrast enhanced MRI percent enhancement time course. The experiment also included the slice location to capture how for a slice was from the tumor edge. From these inputs, the CNN predicted a net proliferation rate, k, for each slice, which was then used with Eq. (8) to predict cellularity at V2 and V3. Thus, the CNN had three outputs: k (to be compared to the calibrated k from the CBBM), and at t = V2 and V3 (to be compared to measured data). It is noted that the CNN did not directly predict A Eq. (8) to obtain it.

[0130] CNN architecture and training. The study used an encoder-decoder architecture following a U-Net architecture (Figure 10, panel A). Holding a random 20% of patients us a testing set, the experiment trained on 80% of patients in a five-fold cross validation. The loss function used in training the network was visualized in Figure 10, panel B. Results from this method are referred to as “CNN predictions’*. [0131] Statistical methods. For each patient in the testing cohort, the experiment calculated and evaluated (see Table 2) total tumor cellularity (TTC), total tumor volume (TTV), and voxel-wise percent change in cellularity (ATC(x)). The CCC between modeled (i.e., tiie CBBM or CNN prediction) and measured TTC or TTV was calculated across die patient cohort, while the median tod interquartile range (interquartile range) of the absolute difference between modeled (again, the CBBM or CNN prediction) and measured A7 C(x) was calculated per patient. The study also computed the median (interquartile range) of the median differences in voxel-wise differences over the cohort. Receiver operating characteristic (ROC) curve analysis was used to assess the ability to predict pCR status using TTC and TTV at V3 as determined by the measured data, as well as by the CBBM and CNN predictions. For patients who received Taxol, the study reported the area under the ROC curve, sensitivity, and specificity.

Table 2: Summary of metrics used to assess tumor status and response.

[0132] Results

[0133] Patient demographics. After applying the exclusion criteria based on calibration quality, out study populating included 1 18 patients with a median (range) age of 51 (29-78) years. The model was trained using the random fivefold cross validation with 94 patients (51 [29-78] years) and evaluated performance on the 24 (47 [31-62] years) patients reserved for the testing set. Of the testing patients, n - 15 (51 [31 -60] years) received Taxol as their second course of therapy (see Figure 9, panel A for therapy timeline, n - 9 additional patients received targeted therapy) and were used in the ROC analysis to differentiate pCR from residual disease after completion of NAC.

[0134] Model performance. Figure 11 , panel A presents a representative pCR (left column) and non-pCR (right column) patient, (“Representative” meaning the patient with the median absolute error in the CNN predicted TTC at V3.) hi panels A and B, the study plotted the measured, CBBM, and CNN predicted voxel-wise cellularity for the center tumor slice at all force visits. For the pCR patient (Panel A), the median (interquartile range) of the CBBM ( ) ( ) p y

[0135] To assess error over the whole tumor, the experiment analyzed the TTV and TTC. Figure 11, panels C and D show three-dimensional renderings of the measured, CBBM, and CNN predicted tumor volumes for the pCR and non-pCR patient, respectively. The pCR patient has measured and CBBM TTV at V3 of 0.0 cm⁵ and 0.0 cm³, respectively. The CNN predicted a tumor volume at V3 of 0.9 (0.86-1.9) cm³. (All CNN predictions of TTV and TTC are reported as median (range) across the cross-validation predictions.) The non-pCR patient had measured and CBBM: tumor volumes at V3 of 6.9 cm³ and 6.7 cm³, respectively,

[0136] Figure 12, panel A shows the change in TTC from V 1 to V3 for the CBBM versus the measured data. (Each point represents one patient.) Panel B shows the same metric for the CNN prediction versus the measured data. Each point shows the median and range across the five-fold cross validation. The study obtained CCC values from these plots of 0.98 and 0.95 for the CBBM and CNN prediction, respectively. Repeating this analysis for TTV, yielded CCC values of 0.97 and 0.94, respectively.

[0137] Figure 12, panels C and D present the median (interquartile range) of the absolute cohort, the median (interquartile range) of the median absolute differences reveals 0.0% (0.0%-21%) for the CBBM and 0.0% (0.0%-38%) for the CNN prediction. Table 3 summarizes all metrics presorted in Figure 12 across the cohort for V2 and V3.

Table 3: Statistical analysis of metrics across testing cohort.

[0138] For die patients (n ~ 15) who received Taxol as their second course of therapy, the study used ROC analysis to evaluate the ability to differentiate between pCR from residual disease after NAC. The ROC analysis using TTC at V3 returned mean (95% confidence interval) area under curve values of 0.89 (0.58-1.0), 0.86 (0.36-1.0), and 0.72 (0.34-0.94) for the measured data, CBBM, and CNN predictions, respectively. Repeating this analysis using TTV at V3, the study obtained area under curve values of 0.88 (0.63-1.0), 0.86 (0.36-1.0), and 0.72 (0.40-0.95) for the measured data, CBBM, and CNN predictions, respectively. Sensitivity, and specificity based on the optimal cut point are shown in Table 4.

Table 4: ROC analysis. ’CBBM refers to the calibrated, biology-based mathematical model.

² AUG refers to area under the ROC curve.

[6139] Discussion

[0140] Ibis study linked a biology-based mathematical model with deep learning to allow for the prediction of biologically interpretable model parameters describing the spatiotemporal dynamics of TNBC tumor response to NAC using only MRI data acquired before therapy initiation. The study reported a CCC of 0.95 between the CNN prediction and measured data for change in TTC from VI to V3, and 0.98 between the CBBM and measured data for the same prediction. While the CCC for the CNN prediction is lower than that of the CBBM, it is sufficiently high (i.e., > 0.9) support the determination that pre-treatment imaging data contains enough information to predict the response of TNBC to NAC. Further support to this is provided when considering the TTV results with CCC values of 0.94 and 0.97 for the CNN prediction and CBBM, respectively. These results were compared to those in Jarrett et al. (2022), which uses a subset of the ARTEMIS dataset (n ~ 56) to calibrate a more comprehensive, biology-based mathematical model to the V 1 and V2 MRI data, then predict to V3. In particular, Jarrett et al. found CCCs of 0.95 and 0.94 for change in TTC and TTV from VI to V3, respectively. While the CNN predictions show high accuracy and comparability to the CBBM global metrics (TTC, TTV), a difference in accuracy is observed 38%) for the CNN predictions. This result likely occurs because the CNN does not perfectly reconstruct the CBBM, and the local metric is more sensitive to errors than the global metrics.

[0141] When using TTC and TTV at V3 obtained from the CNN predicted parameters in the ROC analysis to differentiate between pCR andnon-pCR patients, the study reported area under curve of 0.72 for both TTC and TTV, which supports this determination. Previous deep learning based methods achieved area under curve of 0.77 when using a CNN to directly predict pCR for 33 breast cancer patients from pre-treatment dynamic contrast enhanced MRI, but lack the biological interpretability of the methodology. The CBBM attained an area under curve of 0.86 for both TTC and TTV, which is comparable to the area under curve of 0.89 for the calibrated biology-based model in Jarrett et al. While these methods using longitudinal measurement yield higher area under curve values than the CNN, it is at the expense of requiring mid-treatment imaging data, ft is noted that the ROC analysis covered a smaller patient cohort (n = 15) than these previous two studies, which may limit the ability to directly compare values between the studies.

[0142] Several possible modification to the methods used in this study are also envisions. First, as with deep learning-based methods, the methods use a training cohort, therefore die performance of the model can depend on the quantity and quality of data. The ARTEMIS trial dataset provides a large cohort of patients who received uniform, high-quality imaging at the same site with the same quality assurance/quality control specifications. While other large datasets exist, they are often from multi-site studies, which necessarily increases heterogeneity in the training cohort.

[0143] Although the present experiment utilized the biology-based model presented in Eq. (8) for computational efficacy, it is also envisioned that more complex models can be utilized to further provide accurate predictive results. For example, the biological-based model can additionally include a drug term, such as that described in Eq. (1), and/or utilize a network capable of predicting separate proliferation and drug effects.

[0144] In conclusion, the present study shows the integration of a biology-based mathematical model wife a convolutional neural network to provide spatio-temporally resolved predictions of TNBC response to MAC using only pretreatment MR! data. By combining biology- and data-based methods, the models were able to attain high predictive accuracy before initiating therapy while maintaining the interpretability of the prediction afforded by fee biology-based model calibrated to pre- and on-treatment images.

[0145] Other Machine Learning. In addition to the machine learning features described above, fee various analysis system can be implemented using one or more artificial intelligence and machine learning operations. The term “artificial intelligence” can include any technique that enables one or more computing devices or comping systems (i.e., a machine) to mimic human intelligence. Artificial intelligence (Al) includes but is not limited to knowledge bases, machine learning, representation learning, and deep learning. The term “machine learning” is defined herein to be a subset of Al feat enables a machine to acquire knowledge by extracting patterns from raw data, Machine learning techniques include, but are not limited to, logistic regression, support vector machines (SVMs), decision trees, Naive Bayes classifiers, and artificial neural networks. The term “representation learning” is defined herein to be a subset of machine learning that enables a machine to automatically discover representations needed for feature detection, prediction, or classification from raw data. Representation learning techniques include, but are not limited to, autoencoders and embeddings. The term “deep learning” is defined herein to be a subset of machine learning that enables a machine to automatically discover representations needed for feature detection, prediction, classification, etc., using layers of processing. Deep learning techniques include but are not limited to artificial neural networks or multilayer perceptron (MLP).

[0146] Machine learning models include supervised, semi-supervised, and unsupervised learning models. In a supervised learning model, the model leams a function that maps an input (also known as feature or features) to an output (also known as target) during training wife a labeled data set (or dataset), In an unsupervised learning model, the algorithm discovers patterns among data. In a semi-supervised model, the model leams a function that maps an input (also known as a feature or features) to an output (also known as a target) during training with both labeled and unlabeled data.

[0147] Neural Networks. An artificial neural network (ANN) is a computing system including a plurality of interconnected neurons (e.g., also referred to as ‘‘nodes”). This disclosure contemplates that the nodes can be implemented using a computing device (e.g., a processing unit and memory as described herein). The nodes can be arranged in a plurality of layers such as an input layer, an output layer, and optionally one or more hidden layers with different activation functions. An ANN having hidden layers can be referred to as a deep neural network or multilayer perceptron (MLP). Each node is connected to one or more other nodes in the ANN. For example, each layer is made of a plurality of nodes, where each node is connected to all nodes in the previous layer. The nodes in a given layer are not interconnected with one another, i.e., the nodes in a given layer function independently of one another. As used herein, nodes in the input layer receive data from outside of the ANN, nodes in fee hidden layers) modify the data between the input and output layers, and nodes in fee output layer provide the results. Each node is configured to receive an input, implement an activation function (e.g., binary step, linear, sigmoid, tanh, or rectified linear unit (ReLU)), and provide an output in accordance with the activation function. Additionally, each node is associated with a respective weight. ANNs ate trained with a dataset to maximize or minimize an objective function. In some implementations, the objective function is a cost function, which is a measure of the ANN’S performance (e.g., error such as LI or L2 loss) during training, and the training algorithm tunes the node weights and/or bias to minimize the cost function. This disclosure contemplates feat any algorithm that finds the maximum or minimum of fee objective function can be used for training the ANN. Training algorithms for ANNs include but are not limited to backpropagation. It should be understood that an ANN is provided only as an example machine learning model. This disclosure contemplates that the machine learning model can be any supervised learning model, semi-supervised learning model, or unsupervised learning model. Optionally, the machine learning model is a deep learning model Machine learning models ate known in the art and are therefore not described in further detail herein.

[0148] A convolutional neural network (CNN) is a type of deep neural network that has been applied, for example, to image analysis applications. Unlike traditional neural networks, each layer in a CNN has a plurality of nodes arranged in three dimensions (width, height, depth). CNNs can include different types of layers, e.g., convolutional, pooling, and folly- connected (also referred to herein as “dense”) layers. A convolutional layer includes a set of filters and performs the bulk of the computations. A pooling layer is optionally inserted between convolutional layers to reduce the computational power and/or control overfitting (e.g., by downsampling). A folly-connected layer includes neurons, where each neuron is connected to all of the neurons in the previous layer. The layers ate stacked similarly to traditional neural networks. GCNNs ate CNNs that have been adapted to work on structured datasets such as graphs.

[0149] As used herein, the term “encoder-decoder Al model’* describes an architecture of a machine learning system in which input data are encoded or coded in order then to be decoded again immediately afterward. In the middle between the encoder and the decoder the necessary data are present as a type of feature vector. During decoding, depending on the training of the machine learning model, specific features in the input data can then be specially highlighted.

[0150] As used herein, the term “U-net” describes an architecture of a machine learning system which is based on a convolutional network architecture. This architecture is particularly well suited to a fest and accurate segmentation of digital images in the biological/medical field. A further advantage of such a machine learning system is that it manages with fewer training data and allows a comparatively accurate segmentation.

[0151] Other Supervised/. earning Models. A logistic regression (LR) classifier is a supervised classification model that uses the logistic function to predict the probability of a target, which can be used for classification. LR classifiers are trained with a data set (also referred to herein as a “dataset’*) to maximize or minimize an objective function, for example, a measure of the LR classifier’s performance (e.g., error such as LI or L2 loss), daring training. This disclosure contemplates that any algorithm that finds the minimum of the cost function can be used. LR classifiers are known in the art and are therefore not described in further detail herein.

[0152] A Naive Bayes’ (NB) classifier is a supervised classification model that is based on Bayes’ Theorem, which assumes independence among features (i.e„ the presence of one feature in a class is unrelated to the presence of any other features). NB classifiers are trained with a data set by computing the conditional probability distribution of each feature given a label and applying Bayes’ Theorem to compute the conditional probability distribution of a label given an observation. NB classifiers are known in the art and are therefore not described in further detail herein.

[0153] A k-NN classifier is an unsupervised classification model that classifies new data points based on similarity measures (e.g., distance functions). The k-NN classifiers are trained with a data set (also referred to herein as a “dataset”) to maximize or minimize a measure of the k-NN classifier’s performance during training. This disclosure contemplates any algorithm that finds the maximum or minimum. The k-NN classifiers ate known in the art and are therefore not described in further detail herein.

[0154] Conclusion

[0155] The methods, systems, and devices discussed above are examples. Various embodiments may omit, substitute, or add various procedures or components as appropriate. For instance, in alternative configurations, the methods described may be performed in an order different from that described, and/or various stages may be added, omitted, and/or combined. Also, features described with respect to certain embodiments may be combined in various other embodiments. Different aspects and elements of the embodiments may be combined in a similar manner. Also, technology evolves and, thus, many of the elements are examples that do not limit the scope of the disclosure to those specific examples.

[0156] The foregoing method descriptions and the process flow diagrams are provided merely as illustrative examples and are not intended to require or imply that the operations of various embodiments must be performed in the order presented. As will be appreciated by one of skill in the ait the order of operations in the foregoing embodiments may be performed in any order. Words such as “thereafter," “then," “next,” etc. are not intended to limit the order of the operations; these words are simply used to guide the reader through the description of the methods. Further, atty reference to claim elements in die singular, for example, using the articles “a,” “an” or “the” is not to be construed as limiting the element to the singular.

[0157] While the terms “first” and “second” are used herein to describe data transmission associated with a subscription and data receiving associated with a different subscription, such identifiers are merely for convenience and are not meant to limit various embodiments to a particular order, sequence, type of network or carrier, [0158] Various illustrative logical blocks, modules, circuits, and algorithm operations described in connection with the embodiments disclosed herein may be implemented as electronic hardware, computer software, or combinations of both. To clearly illustrate this interchangeability of hardware and software, various illustrative components, blocks, modules, circuits, and operations have been described above generally in terms of their functionality. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the overall system. Skilled artisans may implement the described functionality in varying ways for each particular application, but such embodiment decisions should not be interpreted as causing a departure from the scope of the claims.

[0159] The hardware used to implement various illustrative logics, logical blocks, modules, and circuits described in connection with the embodiments disclosed herein may be implemented or performed with a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the ftmctions described herein. A general-purpose processor may be a microprocessor, but in the alternative, the processor may be any conventional processor, controller, microcontroller, or state machine. A processor may also be implemented as a combination of computing systems (e.g., a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors in conjunction with a DSP core, or any other such configuration. Alternatively, some operations or methods may be performed by circuitry that is specific to a given function.

[0160] In one or more example embodiments, the functions described may be implemented in hardware, software, firmware, or any combination thereof. If implemented in software, the ftmctions may be stored as one or more instructions or codes on a non-transitory computer-readable medium or non-transitory processor-readable medium. The operations of a method or algorithm disclosed herein may be embodied in a processor-executable software module, which may reside on a non-transitory computer-readable or processor-readable storage medium. Non-transitory computer-readable or processor-readable storage media may be any storage media that may be accessed by a computer or a processor. By way of example but not limitation, such non-transitory computer-readable or processor-readable media may include RAM, ROM, EEPROM, FLASH memory, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium that may be used to store desired program code in the form of instructions or data structures and that may be accessed by a computer. Disk and disc, as used herein, includes compact disc (CD), laser disc, optical disc, digital versatile disc (DVD), floppy disk, and Blu-ray disc, where disks usually reproduce data magnetically, while discs reproduce data optically with lasers. Combinations of the above are also included within the scope of non-transitory Computer- readable and processor-readable media. Additionally, the operations of a method or algorithm may reside as one or any combination or set of codes and/or instructions on a non- transitory processor-readable medium and/or computer-readable medium, which may be incorporated into a computer program product.

[0161] Those of skill in the art will appreciate that information and signals used to communicate the messages described herein may be represented using any of a variety of different technologies and techniques. For example, data, instructions, commands, information, signals, bits, symbols, and chips that may be referenced throughout the above description may be represented by voltages, currents, electromagnetic waves, magnetic fields or particles, optical fields or particles, or any combination thereof.

[0162] Terms “and” and “or” as used herein may include a variety of meanings that also are expected to depend at least in part upon the context in which such terms are used.

Typically, “or” if used to associate a list, such as A, B, or C, is intended to mean A, B, and C, here used in the inclusive sense, as well as A, B, or C, here used in the exclusive sense. In addition, the term “one or more,” as used herein, may be used to describe any feature, structure, or characteristic in the singular or may be used to describe some combination of features, structures, or characteristics. However, it should be noted that this is merely an illustrative example, and claimed subject matter is not limited to this example. Furthermore, the term “at least one of if used to associate a list, such as A, B, or C, can be interpreted to mean any combination of A, B, and/or C, such as A, AB, AC, BC, AA, ABC, AAB, AABBCCC, and the like. [0163] As used herein, the term “is associated with*’, as in A is associated with B, means that A refers to B, is B, identifies a feature of B, or indicates that B exists. For example, an image that is associated with a biological state can, by virtue of the data it contains, refer to that biological state, indicate the presence of that biological state, identify a feature of that biological state, or simply indicate feat feat biological state exists.

[0164] Further, while certain embodiments have been described using a particular combination of hardware and software, it should be recognized that other combinations of hardware and software are also possible. Certain embodiments may be implemented only in hardware, or only in software, or using combinations thereof. In one example, fee software may be implemented with a computer program product containing computer program code or instructions executable by one or more processors for performing any or all of the steps, operations, or processes described in this disclosure, where the computer program may be stored on a non-transitory computer-readable medium. The various processes described herein can be implemented on fee same processor or different processors in any combination. [0165] Where devices, systems, components or modules are described as being configured to perform certain operations or functions, such configuration can be accomplished, for example, by designing electronic circuits to perform fee operation, by programming programmable electronic circuits (such as microprocessors) to perform the operation such as by executing computer instructions or code, or processors or cores programmed to execute code or instructions stored on a non-transitory memory medium, or any combination thereof. Processes can communicate using a variety of techniques, including, but not limited to, conventional techniques for inter-process communications, and different pairs of processes may use different techniques, or fee same pair of processes may use different techniques at different times.

[0166] While the disclosure has been described in detail wife reference to exemplary embodiments, those skilled in fee art will appreciate that various modifications and substitutions may be made thereto without departing from the spirit and scope of the disclosure as set forth in the appended claims. For example, elements and/or features of different exemplary embodiments may be combined with each other and/or substituted for each other within the scope of this disclosure and appended claims.

[0167] The disclosures of each and every publication cited herein are hereby incorporated by reference in their entirety.

Reference List #1 [1] Li W, Arasu V, Newitt DC, Jones EF, Wilmes L, Gibbs J, et al. Effect of MR imaging contrast thresholds on ptediction of neoadjuvant chemotherapy response in breast cancer subtypes: a subgroup analysis of the ACRIN 6657/l-SPY 1 TRIAL Tomography. 2016,2 (4): 378-87.

[2] Musalt BC, Abdelhafez AH, Adrada BE, Candelaria RP, Mohamed RM, Boge M, et al Functional tumor volume by fest dynamic contrast-enhanced MRI for predicting neoadjuvant systemic therapy response in triple-negative breast cancer. Journal of Magnetic Resonance Imaging. 2021;54(l):25l-60.

[33 Wu C, Jarrett AM, Zhou Z, Elshafeey N, Adrada BE, Candelaria RP, et al. MRI-based digital models forecast patient-specific treatment responses to neoadjuvant chemotherapy in triple-negative breast cancer. Cancer Research. 2022;82(18):3394- 404.

[4] Ronneberger O, Fischer P, Brox T. U-net: Convolutional networks for biomedical image segmentation. InMedical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part HI 182015:234- 241.

[5] He K, Zhang X, Ren S, Sun J. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. InProceedings of the IEEE international conference on computer vision 2015:1026-1034.

[6] Hormuth, D.A., Eldridge, S.L., Weis, J.A., Miga, M I. and Yankeelov, T.E., 2018. Mechanically coupled reaction-diffusion model to predict glioma growth: methodological details. Cancer Systems Biology: Methods and Protocols, pp.225-241.

17] Deise Raquel Barpe, Daniela Domelies Rosa, and Pedro Eduardo Froehlich.

Pharmacokinetic evaluation of doxorubicin plasma levels in normal and overweight patients with breast cancer and simulation of dose adjustment by different indexes of body mass. Eur J Pharm Sci, 41(3-4):458-463, November 2010.

[8] Hua-Qing Zhu, Ai-ping Hu, Jian Chen, and Peng Shen. Pharmacokinetics and safety of cyclophosphamide and docetaxel in a hemodialysis patient with early stage breast cancer: a case report. BMC Cancer, 15:917, November 2015.

[9] Garth Powis, Phillip Reece, David L. Ahmann, and James N. Ingle. Effect of body weight on the pharmacokinetics of cyclophosphamide in breast cancer patients. Cancer Chemother. Pharmacol., 20(3):219-222, September 1987. [10] Taisuke Mori, Yoshiyuki Kinoshita, Ai Watanabe, Takeshi Yamaguchi, Kenichi Hosokawa, and Hideo Honjo. Retention of paclitaxel in cancer cells for 1 week in vivo and in vitro. Cancer Chemother Pharmacol, 58(5):665-672, November 2006.

[11] Adam Paszke, Sam Gross, Sotimith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Amiga, and Adam Lerer. Automatic differentiation in PyTorch. Long Beach, CA, USA, 2017.

[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving Deep into Rectifiers:

Surpassing Human-Level Performance on ImageNet Classification, February 2015. arXiv: 1502.01852 [cs] version: 1.

[13] ChengyueWu, Angela M. Jarrett, Zijian Zhou, Nabil Elshafeey, Beatrix E. Adrada, Rosalind P. Candelaria, Rania M.M. Mohamed, Medine Boge, Lei Huo, Jason B. White, Debu Tripafoy, Vicente Valero, Jennifer K, Litton, Clinton Yam, Jong Bum Son, Jingfei Ma, Gaiane M. Rauch, and Thomas E. Yankeelov. MRl-Based Digital Models Forecast Patient-Specific Treatment Responses to Neoadjuvant Chemotherapy in Triple-Negative Breast Cancer. Cancer Research, 82(18): 3394- 3404, September 2022,

[14] Casey E. Stowers, Chengyue Wu, Sidharth Kumar, Ernesto A.B.F. Lima, Xhan Zu, Clinton Yam, Jong Bum Son, Jingfei Ma, Jonathan I. Tamir, Gaiane M. Rauch, Thomas E. Yankeelov. Developing MRI-based digital-twins via mathematical modeling and deep learning to predict the response of triple-negative breast cancer to neoadjuvant therapy [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023 ;83(7_Suppl): Abstract nr 845.

[15] PCT/US2022/044285

[16] Stowers CE, Wu C, Xu Z, Kumar S, Yam C, Son JB, Ma J, Tamir JI, Rauch GM, Yankeelov TE. “Using pre-treatment MRI data to drive an integrated mechanism- and data-driven model to predict the response of breast cancer to therapy.” May 2024, SIAM Conference in Imaging Science, Atlanta, GA, USA. (Invited)

[17] Stowers CE, Wu C, Kumar S, Xu Z, Yam C, Son JB, Ma J, Tamir JI, Rauch GM, Yankeelov TE. “Integrating mechanism-based and data-driven modeling to predict the response of triple negative breast cancer to therapy”. Dec 2023, San Antonio Breast Cancer Symposium, San Antonio, TX, USA.

[18] Stowers CE, Wu C, Kumar S, Lima EABF, Xu Z, Ma J, Tamir JI, Rauch GM, YankeelovTE. “Developing MRI-based digital-twins via mathematical modeling and deep learning to predict the response of triple-negative breast cancer to neoadjuvant therapy”. Apr 2023, AAC'B Annual Meeting, Orlando, FL, USA.

[19] Stowers CE, Wu C, Xu Z, Kumar S, Yam C, Son JB, Ma J, Tamir JI, Rauch GM, Yankeelov TE. “Integrating mechanism-based modeling and MRI data to improve predictions of breast cancer response to chemotherapy.” June- July 2024, Joint Annual Meeting of the Korean Society for Mathematical Biology and the Society for Mathematical Biology, Seoul, Republic of Korea.

Reference List #2

[1] Liedtke C, Mazouni C, Hess KR, Andre F, Tordai A, Mejia JA, Symmans WF, Gonzalez-Angulo AM, Hennessy B, Green M, Cristofanilli M, Hortobagyi GN, Pusztai L. Response to Neoadjuvant Therapy and Long-Tenn Survival in Patients With Triple-Negative Breast Cancer. ICO. 2008 Mar 10;26(8):l275-8L

[2] Liu SV, Melstrom L, Yao K, Russell CA, Sener SF. Neoadjuvant therapy for breast cancer. J Surg Oncol. 2010 Mar 15;101(4):283-91.

[3] Pathological Complete Response in Neoadjuvant Treatment of High-Risk Early-Stage Breast Cancer: Use as an Endpoint to Support Accelerated Approval Guidance for Industry. US. Department of Health and Human Services Food and Drug Administration Oncology Center of Excellence Center for Drug Evaluation and Research (CDER) Center for Biologies Evaluation and Research (CBER); 2020.

[4] Spring LM, Fell G, Arfe A, Sharma C. Greenup R, Reynolds KL. Smith BL, Alexander B, Moy B, Isakoff SJ, Parmigiani G, Trippa L, Bardia A. Pathologic Complete Response after Neoadjuvant Chemotherapy and Impact on Breast Cancer Recurrence and Survival: A Comprehensive Metaanalysis. Clinical Cancer Research. 2020 Jun 15 ;26( 12); 2838-48.

[5] Wu K, Yang Q, Liu Y, Wu A, Yang Z; Meta-analysis on the association between pathologic complete response and triple-negative breast cancer after neoadjuvant chemotherapy. World J Surg One. 20l4;12(l):95. [6] Schmid P, Cortes J, Pusztai L, McArthur H, Kummel S, Bergh J, Denkert C, Park YH, Hui R, Harbeck N, Takahashi M, Foukakis T, Fasching PA, Cardoso F, Untch M, Jia L, Karantza V, Zhao J, Aktan G, Dent R, O’Shaughnessy J. Pembrolizumab for Early Triple-Negative Breast Cancer. N Engl J Med. 2020 Feb 27;382(9):81O-21.

[7] Weis JA, Miga MI, Arlinghaus LR, Li X, Abramson V, Chakravarthy AB, Pendyala P, Yankeelov TE, Predicting the Response of Breast Cancer to Neoadjuvant Therapy Using a Mechanically Coupled Reaction Diffusion Model. Cancer Research. 2015 Nov 15;75(22):4697-707.

[8] Jarrett AM, Hormuth DA, Barnes SL, Feng X, Huang W, Yankeelov TE. Incorporating drug delivery into an imaging-driven, mechanics-coupled reaction diffusion model for predicting toe response of breast cancer to neoadjuvant chemotherapy: theory and preliminary clinical results. Phys Med Biol 2018 May 17;63(l0):l05015.

[9] Jarrett AM, Kazerouni AS, Wu C, Virostko J, Sorace AG, DiCarlo JC, Hormuth DA, Ekrut DA, Part D, Goodgame B, Avery S, Yankeelov TE. Quantitative magnetic resonance imaging and tumor forecasting of breast ameer patients in the community setting. Nat Protoc. 2021 Nov;16(ll):5309~38.

[10] Wu C, Jarrett AM, Zhou Z, Elshafeey N, Adrada BE, Candelaria RP, Mohamed RMM, Boge M, Huo L, White JB, Tripathy D, Valero V, Litton JK, Yam C, Son JB Ma J, Rauch GM, Yankeelov TE. MRI-Based Digital Models Forecast Patient- Specific Treatment Responses to Neoadjuvant Chemotherapy in Triple-Negative Breast Cancer. Cancer Research. 2022 Sep 16,82(18):3394-404.

[11] Peng Y, Cheng Z, Gong C, Zheng C, Zhang X, Wu Z, Yang Y, Yang X, Zheng J, Shen J. Pretreatment DCE-MRLBased Deep Learning Outperforms Radtomics Analysis in Predicting Pathologic Complete Response to Neoadjuvant Chemotherapy in Breast Cancer. Front Oncol. 2022 Mar 10; 12:846775.

[12] Cain EH, Saha A, Harowicz MR, Marks JR, Marcom PK, Mazurowski MA. Multivariate machine learning models for prediction of pathologic response to neoadjuvant therapy in breast cancer using MR] features: a study using an independent validation set Breast Cancer Res Treat. 2019 Jan;173(2):455-63.

[13] Moghadas-Dastjerdi H, Rahman SETH, Sannachi L, Wright FC, Gandhi S, Trudeau ME, Sadeghi-Naini A, Czaraota GJ. Prediction of chemotherapy response in breast cancer patients at pretreatment using second derivative texture of CT images and machine learning. Translational Oncology. 2021 Oct; 14(10): 101183.

[14] Saednia K, Lagree A, Alera MA, Fleshner L, Shiner A, Law E. Law B, Dodington DW, Lu FI, Tran WT, Sadeghi-Naini A. Quantitative digital histopathology and machine learning to predict pathological complete response to chemotherapy in breast cancer patients using pre-treatment tumor biopsies. Sd Rep. 2022 Jun 11 ; 12(1 ):9690.

[15] Zhou Z, Adrada BE, Candelaria RP. Elshafeey NA, Boge M, Mohamed RM, Pashapoor S, Sim J, Xu Z, Panthi B, Son JB, Guirguis MS, Patel MM, Whitman GJ, Moseley TW, Scoggins ME, White JB, Litton JK, Valero V, Hunt KK, Tripathy D, Yang W, Wei P, Yam C, Pagel MD, Rauch GM, Ma J. Prediction of pathologic complete response to neoadjuvant systemic therapy in triple negative breast cancer using deep learning on multiparametric MRI. Sci Rep. 2023 Jan 20;l 3(1): 1171.

(16] Qu Y, Zhu H, Cao K, Li X, Ye M, Sun Y. Prediction of pathological complete response to neoadjuvant chemotherapy in breast cancer using a deep learning (DL) method. Thoradc Cancer. 2020 Mar;l l(3):651-8.

[17] Altman DG. Practical statistics for medical research. Boca Raton, Fla: Chapman & Hall/CRC; 1999. 611 p.

[18] Zou KH, O’Malley AJ, Mauri L. Receiver-Operating Characteristic Analysis for Evaluating Diagnostic Tests and Predictive Models. Circulation. 2007 Feb 6;115(5):654-7.

[19] Yam C, Yen EY, Chang JT, Bassett RL, Alatrash G, Garber H, Huo L, Yang F, Philips AV, Ding QQ, Lim B, Ueno NT, Kannan K, Sun X, Sun B, Parra Cuentas ER, Symmans WF, White JB, Ravenberg E, Seth S, Guerriero JL, Rauch GM, Damodaran S, Litton JK, Wargo JA, Hortobagyi GN, Futreal A, Wistuba II, Sun R, Moulder SL, Mittendorf EA. Immune Phenotype and Response to Neoadjuvant Therapy in TripleNegative Breast Cancer. Clinical Cancer Research. 2021 Oct l;27(19):5365-75.

[20] Yam C, Abuhadra N, Sun R, Adrada BE, Ding QQ, White JB, Ravenberg EE, Clayborn AR, Valero V, Tripathy D, Damodaran S, Arun BK, Litton JK, Ueno NT, Murthy RK, Lim B, Baez L, Li X, Buzdar AU, Hortobagyi GN, Thompson AM, Mittendorf EA, Rauch GM, Candelaria RP, Huo L, Moulder SL, Chang JT. Molecular Characterization and Prospective Evaluation of Pathologic Response and Outcomes with Neoadjuvant Therapy in Metaplastic Triple-Negative Breast Cancer. Clinical Cancer Research. 2022 Jul 1 ;28( 13);2878--89. [21] Yankeelov TE, Pickens DR, Price RR, editors. Quantitative MR] in Cancer.

Quantitative MRI in cancer. Boca Raton: CRC Press; 2012. (Imaging in medical diagnosis and therapy).

[22] Ronneberger Q, Fischer P, Brox T. U-Net: Convolutional Networks for Biomedical

Image Segmentation [Internet]. arXiv; 20(5 [cited 2023 Jan 10].

[23] Ravichandran K, Braman N, Janowczyk A, Madabhushi A. A deep learning classifier for prediction of pathological complete response to neoadjuvant chemotherapy from baseline breast DCE-MRI. In: Mori K, Petrick N, editors. Medical Imaging 2018: Computer-Aided Diagnosis [Internet]. Houston, United States: SPIE; 2018 [cited 2023 Oct 18], p. 11.

[24] Wang H, Yee D. I-SPY 2: a Neoadjuvant Adaptive Clinical Trial Designed to Improve Outcomes in High-Risk Breast Cancer. Cuir Breast Cancer Rep. 2019 Dec;l 1(4)303™ 10.

[25] Atuegwu NC, Ariinghaus LR, Li X, Welch EB, Chakravarthy BA, Gore JC, Yankeelov TE. Integration of diffusion-weighted MRI data and a simple mathematical model to predict breast tumor celluiarity during neoadjuvant chemotherapy. Magn Reson Med. 2011 Dec;66(6): 1689-96.

[26] Levenberg K. A method for the solution of certain non-linear problems in least squares. Quart AppI Math. 1944;2(2): 164-8.

[27] Marquardt DW. An Algorithm fbr Least-Squares Estimation of Nonlinear Parameters. Journal of the Society for Industrial and Applied Mathematics. 1963 Jun; 11 (2):431- 41.

[28] Li W, Arasu V, Newitt DC, Jones EF, Wilmes L, Gibbs J, Komak J, Joe BN, Esserman LJ, Hylton NM. Effect of MR Imaging Contrast Thresholds on Prediction of Neoadjuvant Chemotherapy Response in Breast Cancer Subtypes: A Subgroup Analysis of the ACRIN 6657/I-SPY 1 TRIAL. Tomography. 2016 Dec;2(4):378-87.

[29] Musall BC, Abdelhafez AH, Adrada BE, Candelaria RP, Mohamed RMM, Boge M, Le-Petross H, Arribas E, Lane DL, Spak DA, Leung JWT, Hwang KP, Son JB, Eishafeey NA, Mahmoud HS, Wei P, Sun J, Zhang S, White JB, Raveriberg EE, Litton JK, Damodaran S, Thompson AM, Moulder SL, Yang WT, Pagel MD, Rauch GM, Ma J. Functional Tumor Volume by Fast Dynamic Contrast-Enhanced MRI for Predicting Neoadjuvant Systemic Therapy Response in Triple-Negative Breast Cancer. J Magn Reson Imaging. 2021 Jul;54(l):251-60. [30] Paszke A, Gross S, Chintala S, Chanan G, Yang E, DeVito Z, Lin Z, Desmaison A, Antiga L, Lerer A, Automatic differentiation in PyTorch. In Long Beach, CA, USA; 2017.

[31] Kingma DP, Ba J. Adam: A Method for Stochastic Optimization [Internet], arXiv; 2017 [cited 2023 Mar 29].

Claims

CLAIMS What is claimed is:

1. A method for predicting a patient-speci fic therapeutic response to neoadjuvant treatment in a subject having cancer, the method comprising: receiving imaging data associated with one or more imaging biomarkers of tumor cells in the subject, the imaging data being acquired prior to any neoadjuvant treatment therapy of a cancer; determining model parameters of the tumor cells from the imaging data using a trained artificial intelligence (Al) model; and determining one or more predictive tumor metrics using a biology-based mathematical model, wherein the biology-based mathematical model is configured to evaluate a spatiotemporal growth of the tumor cells in the subject using the model parameters of the tumor cells; wherein the one or more predictive tumor metrics are subsequently used to determine a predicted pathological complete response (pCR.) to adjust or direct neoadjuvant therapy in the subject.

2. The method of claim 1 , wherein the one or more predictive tumor metrics are subsequently used to direct a course of neoadjuvant treatment in the subject.

3.The method of claim 1 , wherein the one or more predictive tumor metrics are subsequently used to compare relative efficacies between two or more courses of neoadjuvant treatment, to subsequently direct neoadjuvant therapy in the subject

4.The method of any one of claims 1-3, wherein the imaging data comprises data obtained from a magnetic resonance imaging (MRI) system.

5. The method of any one of claims 1-4, wherein the imaging data comprises diffusion weighted imaging (DWI) data acquired from an MRI scanner.

6.The method of any one of claims 1-5, wherein the imaging data comprises one or more dynamic contrast enhanced (DCE) MRI.

7. The method of any one of claims 1 -6, wherein the imaging data comprises an apparent diffusion coefficient (ADC) map.

8. The method of any one of claims 1 -7, wherein the imaging data comprises one or more /'/-weighted images and/or /^-weighted images.

9. The method of any one of claims 1-8, wherein the trained Al model determines the one or more predictive tumor metrics using both imaging data and clinical date from die subject.

10. The method of any one of claims 1-9, wherein the trained Al model determines the one or more predictive tumor metrics using both imaging data and genomic data from the subject

11. The method of claim 10, wherein the genomic data includes gene expression levels of a genetic biomarker corresponding to a clinically relevant effect on at least one model parameter of the tumor cells.

12. The method ofany one of claims 10-11, wherein the genomic data is provided as a local input at each of a plurality of splits of the Al model.

13. The method of any one of claims 10-11, wherein the genomic data is provided as a global input to the Al model.

14. The method of any one of claims 1-13, further comprising spatially aligning two or more images of the imaging data at each of a plurality of voxels prior to determining the model parameters using &e integration of the biology-based model with the Al model

15. The method of claim 14, wherein the Al model generates a spatially localized coefficient of the model parameters at each of the plurality of voxels.

16. The method of any one of claims 1-15, wherein the trained Al model comprises an encoder-decoder architecture that takes as input one or more of the imaging data and returns one or more of tumor proliferation rate and predictive tumor metrics.

17. The method of any one of claims 1-16, wherein the Al model comprises a U-Net architecture.

18. The method of any one of claims 1-17, wherein the biology-based mathematical mode! comprises a reaction-diffusion model.

19. The method of any one of claims 1-18, wherein the biology-based mathematical model is further configured to evaluate an effectiveness of a course neoadjuvant treatment.

20. The method of any one of claims 1-19, wherein die cancer is prostate cancer, breast cancer, brain cancer, ovarian cancer, head and neck cancer, pancreatic cancer, cervical cancer, rectal cancer, esophagus cancer, liver cancer, stomach cancer, testicular cancer, vaginal cancer, uterine cancer, vulvar cancer, paranasal cancer, oropharyngeal cancer, or laryngeal cancer.

21. The method of any one of claims 1 -20, wherein the cancer is triple negative breast cancer (TNBC).

22. A system comprising: one or more processors; and a memory having instructions stored thereon, wherein the instructions, when executed by the one or more processors, cause the one or more processors to perform any one of the methods of claims 1-21.

23. The system of claim 22 further comprising an MRI scanner.

24. A noil-transitory computer readable medium having instructions stored thereon, wherein the instructions, when executed by one or more processors, cause the one or more processors to perform and one of the methods of claims 1-21.