WO2023139570A1

WO2023139570A1 - System and method for characterising lung tumours (solid, part-solid and ground-glass) based on invasion criteria by means of pixel distancing and deep learning algorithms

Info

Publication number: WO2023139570A1
Application number: PCT/IB2023/052660
Authority: WO
Inventors: Andres Javier Anaya Isaza
Original assignee: Indigo Technologies S.A.S.
Priority date: 2022-01-21
Filing date: 2023-03-18
Publication date: 2023-07-27
Also published as: CO2022000507A1; WO2023139570A4

Abstract

The present invention relates to systems and methods for characterising lung tumours (solid, part-solid and ground-glass) based on criteria of invasiveness by means of pixel distancing and deep learning algorithms in computerised axial tomography (CAT) scans. The system is made up of an image forming device, an input/output interface, at least one processor, a database, and a memory comprising executable instructions to implement a method comprising: i. Receiving computerised axial tomography images of a subject's chest that show lung lesions, nodules or tumours; ii. Determining the volume of the tumour structure from the chest images of the region of interest; iii. Segmenting a representation of voxels by means of a learning algorithm; iv. Determining a projected final volumetry; v. Calculating distance maps by means of a second unsupervised classification algorithm; vi. Creating a matrix of characteristics from the nodule or tumour structures with a distance vector; vii. Determining a biomarker of invasiveness; and viii. Determining the nodule risk.

Description

SYSTEM AND METHOD FOR CHARACTERIZING LUNG TUMORS (SOLID, SUBSOLI D AND GROUND GLASS) BASED ON I NVASIVE CRITERIA THROUGH PI XELAR DISTANCE AND DEEP LEARNING ALGORI TMOS

TECHNOLOGY SECTOR

The present invention relates to systems and methods for characterizing lung tumors. Particularly, the present invention refers to systems and methods for characterizing lung tumors (solid, subsolid and ground glass) based on invasive criteria using pixel distance and deep learning algorithms in computed axial tomography (CT scans), with application in the field of bioengineering since it applies physical-mathematical concepts and methods to solve life science problems, using analytical and synthetic engineering methodologies.

BACKGROUND OF THE I NVENTION

The World Health Organization (WHO) in its 2018 report, affirms that cancer in general is the pathology responsible for 9.6 million deaths per year, being the second cause of death in the world, where its prevalence is outlined in low- and middle-income countries, with 70% of the manifestation of said disease. The foregoing exposes alarming figures within the group of diseases with the highest lethality, where lung cancer is the most aggressive pathology and the main cause of deaths related to cancer deaths, with figures of 135,000 deaths in countries like the United States in 2020, as well as evidence of new diagnoses of close to 228,000 cases with a high mortality rate, with prognoses of less than 5 years of survival. The foregoing shows the importance of carrying out processes in early detection, with the motivation of mitigating the metastatic presence of lesions or the excessive advance in the rate of tumor growth.

This is how, in the last three decades of the 20th century, as well as the first two decades of the 21st century, computerized axial tomography (CT or CT in English) , has shown superior performance in the early detection of lung cancer, mitigating the final mortality rate by up to 20%, being able to characterize severity through pixel intensity. Then, with the use of the Hounsfiel scale, which defines which are the units of fluid tissue and bone, where they can characterize the morphology of the potentially malignant tissue or nodule, showing that the probability of malignancy is given by factors that correlate the rate of growth and morphology (solid, subsolid and ground glass), as well as geometric characteristics.

Despite the efforts in the characterization under diagnostic imaging methodologies through CT, said image processes are given to human interpretations, of which, notable biases are evident due to the presence of subjective factors and validation processes where it is difficult to reach consensus. In response to the previous challenges, the advent of new artificial intelligence through machine learning and deep learning algorithms allows new methodologies such as (CAD - Computer Aided Detection or Computer Aided Systems), allow the mitigation of the diagnosis product of human error. The aforementioned methods have demonstrated their potential in the replication or simulation of intelligent behavior, in tasks such as classification of pathologies, segmentation of findings, characterization of regions of interest, and elastic recording in prospective pathological prognosis.

For lung cancer detection alone, we can find a large number of investigations and reviews that highlight the functioning and performance of CAD systems. Most of these systems have several stages to reach the diagnosis and generate this “second opinion”. However, all of them have two fundamental elements: (1) detection of nodules (including elimination of false positives) and (2) classification of nodules. Although these two elements already represent a challenge on their own and, despite constant research, today, there is no defined consensus to achieve these objectives. For example, Perez et al. in “Automated lung cancer diagnosis using three-dimensional convolutional neural networks. Med Biol Eng Comput 2020; 58: 1803-15”, implement a series of image preprocessing techniques to segment and obtain nodule candidates. These are generated from the regional maximums after a process of filtering, segmentation and various morphological operations. The generated and real nodes are used to train a 12-layer CNN to eliminate false positives. Finally, these are used to train a second CNN in order to achieve node type prediction. Perez et al. in its development it relies on 3D networks to achieve the last two tasks. On the other hand, Zheng et al. in “Automatic Pulmonary Nodule Detection in CT Scans Using Convolutional Neural Networks Based on Maximum Intensity Projection. IEEE Trans Med Imaging 2020; 39:797-805" avoid using 3D networks in the generation of candidate nodules. Instead, he implements a 2D CNN network on MIP (Maximum intensity projection) images generated with sequential slices of computed tomography. The technique allows him to clearly discriminate between nodules and vessels through a network with a UNet-like architecture.

Despite the effectiveness of the UNet, some of the recently presented research is aimed at exploring or improving new architectures. For example, an architecture based on the V-Net network for segmentation. The network is made up of three pairs of encoder-decoder blocks with residual components. The architecture allows addressing the segmentation problem in 3D, a process that is limited with UNet.

Whether they are segmentation or nodule detection processes, many of them are subject to identifying regions that do not correspond to real nodules (false positives). Consequently, the elimination of false positives is a process inherent to the detection of nodules. Most of the investigations use convolutional networks with complete connections, to generate the classification of the node, that is, if this is a false positive or not.

With the clear identification of the candidate nodules, the investigations focus on discriminating the character of the nodule, that is, if they are potential malignant nodules or not. In this case, the research focuses on single path networks or networks as double or triple paths, as indicated by Apostolopoulos et al. “Experimenting with Convolutional Neural Network architectures for the automatic characterization of Solitary Pulmonary Nodules' malignancy rating. ArXiv 2020” performs a comparison of various architectures where it determines that node detection is more efficient with single path networks.

In the state of the art there are also some patents that have taken the images to perform the diagnosis of diseases, for example the application US2017/046839 provides systems and methods that use a hierarchical analysis framework to identify and quantify biological properties / analytes from imaging data and then identify and characterize one or more medical conditions based on the quantified biological properties / analytes. In some embodiments, the systems and methods incorporate computerized image analysis and data fusion algorithms with patient blood biomarker and clinical chemistry data to provide a multifactorial panel that can be used to distinguish between different disease subtypes, which makes it complex since it requires not only software and hardware elements but patient blood biomarkers and clinical chemistry which are additional processes to a computer-implemented method.

Application US2016/0196648 presents exemplary methods and apparatus that distinguish invasive adenocarcinoma (IA) from non-invasive adenocarcinoma (AI S) in situ depicted in a three-dimensional (3D) computed tomography (CT) image of a region of tissue demonstrating cancerous pathology. The 3D CT image may include representations of ground glass nodules (GGN). The exemplary methods and apparatus perform a 3D histology reconstruction of histology sections of a GGN and record the 3D histology reconstruction with a 3D CT image of the GGN, combining or "fusing" the 3D histology reconstruction with the 3D CT image. Example methods and apparatus map AI regions detected in the 3D histology reconstruction to the 3D CT image and extract texture or shape features associated with AI of the 3D CT image, however, this development does not allow evidencing morphological features of texture or determining the estimation of pixel intensity that allows guaranteeing all the intermediate movements of image processing and its transformation from one state to another.

For its part, the application WO2019157214 describes systems, methods, devices and means to perform medical diagnoses of diseases and conditions using artificial intelligence or machine learning approaches. Deep learning algorithms enable automated analysis of medical images, such as X-rays, to generate predictions with accuracy comparable to that of clinical experts for various diseases and conditions, including those affecting the lung, such as pneumonia. In some embodiments, the transfer learning method further comprises training the pretrained model using a medical image set that is smaller than the large image data set. In some embodiments, the method further comprises making a medical treatment recommendation based on the determination. In some embodiments, the medical image of the lung is a chest X-ray. In some embodiments, the lung disease or disorder is selected from the group consisting of: pneumonia, childhood pneumonia, emphysema, tuberculosis, and lung cancer.

Previous technologies have thought of solving this problem of characterizing nodes in terms of a probability map, on purely two-dimensional representations. The difficulty of this type of approach is the large proportion of details that are left aside, by only delimiting to purely arbitrary classification processes, from which a decision is made that enters into a dichotomy with the diagnostic method of a human radiologist, establishing a statement that leaves aside the invasiveness of a tumor or nodule, especially in its early stages. Due to the above, the need to imitate this intelligent behavior was identified, derived from the Fleischner atlas, where current engineering somewhat ignores the exhaustive rules in terms of differentiating a tumor, under the characteristics closest to reality, as indicated by Fleischner. In this way, the probabilities They are transformed into maps of distances grouped by pixels as a new space of characteristics, finally defining the differential borders of the degree of invasiveness of those early nodular structures, solving the problem of non-differentiability and obtaining a more reliable and less arbitrary biomarker, to predict possible scenarios of growth rate and morphology in density and, above all, invasiveness in situ.

According to the above, the methods used in bioengineering to deal with the recognition of nodule findings focus on the description of broad features in superficial classification of tumor or nodular structures, and poorly correlates with pathology, therefore, there is a need to develop a method and a system that allows correlating the properties of invasive causality and carrying out pathological characterization from the pre-invasive adenocarcinoma phase to invasiveness itself. tu, as well as include other early diagnosis standards that add new feature vectors, from the point of view of artificial intelligence, that include multidimensional representation methods that are capable of solving the early risk thresholding problem and saving lives.

BRIEF DESCRIPTION OF THE I NVENTION

The method and system for characterizing lung tumors (solid, subsolid and ground glass) based on invasive criteria by means of pixel distance and deep learning algorithms in CT scans of the invention presents the analysis of densities and morphological features of the nodules, as a response to the calculation of estimates on the invasive factor, since current developments lack the mapping of this important natural characteristic in the practice of radiology in studies of the thorax, applied to the lung.

Also, it involves the synergy between non-linear learning algorithms such as neural networks with deep learning approaches, which work directly to achieve natural characteristics without external processes of increasing pathological traits, with the difference of achieving a new integration between an unsupervised mechanism by virtue of achieving pixel pairs by grouping or clustering, called pixel distance maps, through methods such as; K-means or Boltzmann or similar machines, presented the characterization of tumor structures and thresholding of a new risk scale based on the invasive factor.

The characterization of distances by group of pixels allows the final representation of the distance map with a fine conformation of the tumor structure, differentiating the invasive factor from the rest of the structure in density to perform early risk thresholding tasks, establishing management protocols in the face of imminent aggressiveness of a nodule or tumor.

This application presents a method to characterize lung tumors based on invasive criteria using pixel distance and deep learning algorithms comprising: i. Receive chest computed tomography images of subjects that show lesions, nodules or lung tumors or any type of anomaly;

Yo. Determine the volume of the tumor structure from the chest images of the region of interest; iii. Segment the voxel representation by means of a learning algorithm; iv. Determine the projected final volumetry; v. Calculate the distance maps using a second unsupervised classification algorithm; saw. Create the training matrix of the structural characteristics of the nodules or tumors and the distance vector; vii. Determine the biomarker of invasiveness; and viii. Determine the risk of the nodule.

Another aspect of the present invention provides a system including an imaging device for imaging a target; at least one processor with a viewer with a file not linked to any CT reconstruction modality or provider (CT scans) in the thorax; a data base; a memory coupled to at least one processor; the memory comprising computer-executable instructions that, when executed by the at least one processor, performs a method comprising: i. Receive chest computed tomography images of subjects showing lung lesions, nodules or tumors or any type of abnormality;

Yo. Determine the volume of the tumor structure from the chest images of the region of interest; iii. Segment the voxel representation by means of a learning algorithm; iv. Determine the projected final volumetry; v. Calculate the distance maps using a second unsupervised classification algorithm; saw. Create the matrix of characteristics of the structures of the nodules or tumors with the distance vector; vii. Determine the biomarker of invasiveness; and viii. Determine the risk of the nodule.

DESCRIPTION OF THE FIGURES

In order to further clarify the above advantages and other features of the present invention, a particular description of the invention will be presented with reference to the specific embodiments thereof which are illustrated in the attached figures. It is appreciated that these figures represent only typical embodiments of the invention and are therefore not considered as limiting in scope.

Figure 1 shows the classification of nodular lesions under parameters: geometric, morphological and topological, as well as their correlation with density factors.

Figure 2 shows the Indira® VNA viewer for a computed tomography in chest studies in lung cancer pathologies. Figure 3 shows steps 1 to 4 of the method to characterize lung tumors (solid, subsolid, and ground glass) based on invasive criteria using pixel distance and deep learning algorithms in CT scans.

Figure 4 shows steps 4 to 7 of the method to characterize lung tumors (solid, subsolid, and ground glass) based on invasive criteria using pixel distance and deep learning algorithms in CT scans.

Figure 5 shows stages 7 and 8 of the method to characterize lung tumors (solid, subsolid, and ground glass) based on invasive criteria using pixel distance and deep learning algorithms in CT scans.

Figure 6 shows the system for characterizing lung tumors (solid, subsolid, and ground glass) based on invasive criteria using pixel distance and deep learning algorithms in CT scans.

DETAILED DESCRIPTION OF THE INVENTION

YO . INTRODUCTION AND DEFINITIONS

The present invention relates to a system and methods to characterize lung tumors (solid, subsolid and ground glass) based on invasive criteria using pixel distance and deep learning algorithms in computed axial tomography (CT scans).

The term “pulmonary nodule” refers to an image of pulmonary or pleural lesions superimposed on normal structures with a more or less similar development in all three dimensions of space. Characterized by presenting an area of increased attenuation, rounded or oval, that does not exceed three centimeters in diameter. This is how, depending on the density Tomographically, these pulmonary nodules can be classified as solid, subsolid, or ground glass.

The term “ground glass nodules” refers to an area of focal attenuation magnification that does not obscure or obscure the underlying vessels.

The term “solid nodules” refers to an area of increased attenuation, due to a collapse of the airspace, which prevents seeing the underlying structures of the normal lung parenchyma.

The term "subsolid nodules" refers to nodules in which there is a mixed appearance, that is, in addition to the ground glass component, there is a variable solid portion.

The term "Invasiveness" refers to the degree of hardness of the tumor structure, given that the more consolidated the group of pixels (greater intensity - evidence of hardness), the higher the correlation of belonging to the category of malignancy of a tumor or nodular structure.

The term "Deep learning" refers to the paradigm of artificial intelligence, which uses neural networks as an active principle and its great contribution is to have a large set of these networks, with many deep layers to maximize their complex behavior and obtain better performance.

The term "CTs" refers to the modality or hardware of Computed Axial Tomography, where it outlines the possibility of scanning the structure of the body under study, through several cross sections of the organ, through x-rays, with the great property of delimiting to a scale that allows differentiating tissue, fluid and bone structures, called Hounsfield units.

The term "Screening" refers to the set of methods or practices that are applied to the population, with the need to study diseases and detect complications in time. The term “LUNG-RADS” refers to the screening process that is carried out on a group of people to detect pathologies oriented to potential lung cancers in time.

The term "Glossary or Fleischner Altas" refers to an atlas of information on good practices and approach to findings, in relation to multiple pathologies oriented to cancers, among others.

The term "predictive biomarker" refers to a biological indicator that gives indications of a pathological event, as well as, when these indicators are manifested, it is possible to use statistical or mathematical approximations to predict their behavior.

The term "Artificial Intelligence System" refers to an application or software that has an algorithm capable of imitating the intelligent behavior of a specific task.

The term “threshold” refers to the possibility of delimiting values to a defined threshold or range.

The term “Adenocarcinoma structures” is a formation of glandular tissue that is located in the internal organs, of which is present in most cancers.

The term "Structures of squamous cell cancers" refers to the classification of non-small cell cancers, which in this case analyzes the lung, where there is its category and probability of occurrence: Squamous cell carcinoma (25% of lung cancers), adenocarcinoma (40% of lung cancers) and large cell carcinoma (10% of lung cancers).

The term "Undifferentiated structures" refers to suspected tumor formations, where their characteristics from the point of view of medical images do not have a clear way of differentiation in terms of the intensity of the object of study in relation to its environment, with edges totally diffuse that hardly differs from other formations or organs in its environment.

“Vendor Neutral Archive (VNA) is an application that stores medical images in a standard format with a standard interface. Therefore, images stored in VNA can be accessed through any workstation, regardless of vendor.

“English Region Of Interest (ROI)” is a volume with intensity levels (gray levels) and positions on the three ordered axes.

The term “Group or Pooling” in the artificial intelligence paradigm refers to the process of grouping pixels in an image.

The term "MaxPool" in the artificial intelligence paradigm refers to obtaining the maximum values of a determined group of pixels.

The term "convolution" refers to a mathematical model with the ability to enhance features of a specific image.

The term "Transposed Convolution" refers to a mathematical model with the ability to enhance the inverse of the characteristics of a specific image, with criteria of edges, contours, curves and contrasts.

The term “Up-conv” in the artificial intelligence paradigm is synonymous with convolution.

The term “Feature Maps” in artificial intelligence is used to refer to an array that stores features of an image, where those features are the set of edges, contours, curves, and other attributes of an image. The term "kernel" takes a small segment of a matrix, to later apply some transformation with one of the previously mentioned methods. It is especially useful to avoid the high computational cost of projecting a small segment of the image and mapping the entire image through this movable window or kernel.

The term “scrolling stride” refers to the movement that maps an image through jumps between pixels with units that can increment every 2, 4, 6, 8, or 1, 3, 5, and 7 pixels.

The term “strides” is the synonym or the way of calling it in English of the previous term called “displacement stride”.

The term “Hadamard Product” is a way of operating matrices and vectors mathematically, so that each term is multiplied 1 to 1 with respect to the other.

The term "RIS-PACS type systems" refers to a system applied to medical imaging units that allows knowing the status of the patient from the moment they enter the hospital unit called RIS, as well as another system that is responsible for storing the images called PACS.

eleven . METHOD FOR CHARACTERIZING LUNG TUMORS

The invention is based on the diagnostic markers that are usually used in clinical practice, through screening atlases such as the LUNG-RADS, where the geometric conformation has a risk threshold meaning, by evidencing those circular, ovoid, lobulated, poly-lobulated and spicular formations, as predictive factors in correlation to the benign or malignant index of a nodular or early tumor structure. Figure 1 shows the classification of nodular lesions under parameters: geometric, morphological and topological, as well as their correlation with density factors, evidencing the classification attributes of the Fleischner glossary. From the above, the categories that correlate the types of nodules such as: solid, subsolid (not solid or frosted glass, partially solid) are evidenced, with the aim of mapping the transformation process of a nodule and evidencing its different changes of state, in relation to the intensity of pixels that evaluate the degree of invasiveness, as a predictive biomarker that can be densely mapped by an artificial intelligence system, through the calculation of the pixel distance per grouping of pixels, in correlation to the invasiveness index of the nodule or tumor, with the aim of estimating or predicting the intermediate movements of the tumor phases, thresholded to risk classes.

The characterization of incidental nodules consists of the dense mapping of pixel distances by clustering, applied to the non-small cell cancer segment, such as: adenocarcinoma structures, squamous cell cancer structures, as well as undifferentiated structures. The foregoing is done with the motivation of recognizing early patterns, where the most reproducible properties of the volumetrics in CTs are exploited, as well as the morphological and geometric features, given that the grouping distances of each category of a tumor structure, implementing a fairly evident and robust decision border, obtaining a new risk threshold based on the interpretation of said group of pixels, through the Fleischner atlas, where current developments in artificial intelligence establish its bases on screening processes such as LUNG-RADS.

The dense mapping of the behavior or active process flow of an incidental tumor is made by means An with the aggressiveness and invasiveness of the tumor. Morphologically, the invasive is the group of pixels with the highest intensity, where clinically this measurement must be obtained and often there are inaccuracies in these, with respect to their early phases. If inaccuracies are obtained, said non-solid or ground glass structures tend to decrease said pattern, quickly transforming into an invasive formation at a growth rate of two to three months, doubling its size and reflecting a higher rate of aggressiveness in correlation to its invasive factor, resulting in a high level of risk through the evidence of a solid component, of which the system establishes a distance vector that will be able to obtain new measurements, providing a clear frontier. crim inant or differential, of the invasiveness index in early stages to make finer predictions, taking into account the degree of aggressiveness implicit in the invasive factor of a nodule or tumor.

Images from CT scans, MRIs, mammograms, and/or chest ultrasounds of subjects that show lesions, nodules or lung tumors or any type of anomaly are stored in a database ( 1 ) where the patient is identified together with his medical history.

In relation to the chest images, these must comply with some initial characteristics of having a projection area of the lung and surrounding soft tissues, without maximum intensity projections or other filters, since this can lead to errors in the next stage.

The images can be viewed with any viewer with a file not linked to any provider (VNA) as shown in Figure 2, where the I NDI RA® viewer is shown, which applies this technology to studies of the thorax in lung oncology pathologies. In this figure, a visual representation with masking segmentation is observed, to make it more evident for the professional in radiology or specialized in oncology.

The same functional principle can be applied to respiratory disorders, where viral pneumonias such as COVI D-19, can have diffuse recognition patterns, such as ground glass, as well as lung consolidation and Crazy Paving, where segmentation tests have been performed on non-differentiable regions.

Therefore, a VNA viewer is one of the most efficient ways to visualize tomographies, since it connects synergistically with the visual channel of specialists, increasing medical-analytical capabilities in complementary diagnostic processes.

Continuing with the description of the stages of the method of the invention, as can be seen in Figure 3, the stage of determining the volume of the tumor structure from the chest images is performed by projecting 3D patches of the region of interest (ROI), which takes voxel representation slices guaranteeing the generalized volume of the tumor, showing a 2D square that will be segmented by a learning algorithm.

In the same Figure 3, it is shown in a preferred embodiment, the stage of segmenting the representation of voxels by means of a learning algorithm, in which a deep learning network algorithm of the encoder-decoder type (Encoder - Decoder) is applied, where the stacking of the convolutional filters allows highlighting the characteristics of the node in relation to its morphology and contour.

The deep learning network algorithm performs a segmentation of regions of interest, such as lesions, tumor tissues or any type of anomaly, which are demarcated by a masking biomarker. In this preferred embodiment, an architecture such as the U-Net network is applied, which is the fundamental base in segmentation networks, since it allows preserving the spatial distribution of the image while extracting its characteristics.

The U-Net network consists of two main elements: an encoder and a decoder. The encoder takes the input image and convolves it, generating increasingly complex feature maps as it goes further into the network. Furthermore, convolutional layers are combined with pooling layers to reduce the size of the maps and thus the computational load.

Likewise, in Figure 3, an example of implementation of stage 3 is observed, where the convolutional neural deep network, which, as indicated, can be the U-net network, has an encoder and a decoder, which are divided into substages. Each substage is made up of a number of convolutional layers before resizing the feature maps via pooling (MaxPool) or transposed convolution (Up-conv). The MaxPool and the Up-conv are part of the encoder and decoder respectively. In the first substep and in the first convolution, the input image can be represented as three matrices or three feature maps, where said maps correspond to the three intensity matrices of the red, green and blue channels (RGB image). Therefore, each j-th feature map (A _j ), generated in the first convolution, would be given by equation ( 1 ) .

Where, is the input image in the Z-th feature map of the M channels that make it up (In this case, M = 3). Ky is the kernel or convolutional filter corresponding to the j-th feature map. The filter is generally 3 x 3 in size and must have the same depth as the input, ie it must have the same number of channels as _X¡ . b _j is the bias added to the convolution of the j-th feature map and f is the activation function of this convolutional layer.

In the second substage of the stage of segmenting the voxel representation by means of a learning algorithm, there is a second convolutional layer, to which the same conditions of the previous substage are applied, however, the input of said layer is the output of the previous layer. The iterative process of the learning algorithm is repeated in a similar way throughout all the convolutions, therefore, the j-th feature map of the Z-th convolutional layer can be represented as shown in equations (2) and (3).

Where A ^(ɩ) are all the feature maps generated in the Z-th layer of the convolutional neural network. So, in the first substage, the network uses a cluster operation after two convolutional operations. The pooling operation (Pooling) is similar to convolutional layers, that is, these layers generate a single value for a window (α _τx , _τy ) that scrolls through the image. The window can be any size and any scroll step. However, the most widely used size is 2x2 with strides of 2. The operation carried out with these parameters reduces the size of the feature maps in half while preserving the total number of these; this process can be governed by equation (4) for each j-th map.

Where a _{Sx Sy} is the window made up of the pixels at the combined positions of 8x and 8y. Also, r and c are the pixel positions that vary up to half the size of the input maps. Consequently, the output of the Z-th clustering layer would be given by the set of all reduced feature maps, as shown in equation (7).

m : Number of feature or channel maps.

Therefore, for the preferred embodiment, the first four substages of the encoder would have the inputs, outputs, and sizes shown in Table 1.

Table 1 . UNet ¹ encoder architecture

In the decoder, the features must be convolved into maps of twice the size. However, in discrete space, convolution reduces the dimensions of the image. That is, suppose you have a single-channel image of size nxn (X e R ^nxn with n even) and convolve it with a filter of size 2 x 2 with strides of 2. The result would generate an output half the size as given in equation (8).

The convolution of equation (8) can be represented as a matrix operation, converting the matrices to vectors and the filter to the sparse matrix version. The equivalent is shown in equation (9), moreover, since the bias b does not affect the dimensions of said operation, it is possible to eliminate it to arrive at the model of the transposed convolution.

In this sense, multiplying by the inverse of the sparse matrix on both sides of the equation, the result of equation (10) is generated.

If the arguments of this result are inverted, that is, if the input image is placed on the left and the output image is placed on the right, we will have an operation equivalent to a convolution, but increasing the dimensions of the image as initially sought. The process is known as transposed convolution or is denoted as Up-Conv, as shown in Figure 3, which shows an upward convolution.

It must be taken into account that, although equation (8) denotes the inverse of a non-square matrix, the process is not carried out since the parameters that make up the filter are unknown, that is, the filter could be replaced by a matrix with the dimensions transposed to the original size and with unknown weights, without having significant repercussions since these would be calculated during training.

After the transposed convolution, the result is concatenated with the previous map copies to the pooling layers and resubmitted to the convolutional layers, as illustrated in Figure 3. The process is repeated the same number of times that were pooled using the Max Pooling function.

Therefore, for the preferred embodiment, the decoder stages would have the inputs, outputs, and sizes shown in Error! Reference source not found..

Table 2. Architecture of the UNet decoder

Then, in stage 4 of the method of the present invention, the final projected volumetry is determined, which comprises segmenting each of the geometric and morphological features of the nodule under study, taking into account the topology of the lesion through masking processes already described, as observed in Figure 4 with the number 4, the total volume is sectioned and then each session is segmented, as well as encoded and decoded, to finally obtain as output the generalized segmented volume of the lesion.

This stage makes it possible to close the domain of the focus of invasiveness, with a view to early detection of clustering pixel intensity, clearly defining the borders of the diffuse edges of partially solid and non-solid nodules.

In the next stage of the method to characterize lung tumors (solid, subsolid and ground glass) based on invasive criteria using pixel distance and deep learning algorithms in CTs, the maps of distances by means of a second unsupervised classification algorithm, see Figure 4 number 5. In this stage, the demarcation is taken as input by means of a masking process, as carried out by the Encoder-Decoder type model (Encoder-Decoder) and a second unsupervised classification algorithm, under the modification of some topological criteria by means of K-means or Boltzmann machines, in which the distance maps are obtained by pairs of pixels that can project the quantum values. criteria that will demarcate the new risk thresholds, where it is delimited to the upper and lower limit of each thresholded measure with respect to the intensity of distances, understood as the new measure of invasiveness that can be projected as a new invasive descriptor of potential adenocarcinoma in situ or invasive.

The automatic segmentation generates an image or a probability map, where the lung nodules can be found. The map is reduced to a binary one considering the pixels with the highest probability as the high binary states and those with the lowest probability as the low states, that is, the ones and zeros of the binary maps. The process is represented mathematically as in equation (1 1 ) .

Multiplying the binary map with the original image, through the Hadamard product, only the pixels of the lung nodule are obtained, that is, each nodule is automatically extracted as a region of interest (ROI). The above operation is represented mathematically as shown in equation (12).

Figure 4 shows how a region of interest (ROI) is defined from the final projected volumetry, that is, a volume with intensity levels (gray levels) and positions on the three ordered axes. Consequently, each voxel that makes up the ROI can be represented as a vector of d dimensions, each dimension being a descriptor or characteristic, as shown in equation (13).

Where j is the subscript associated with each voxel that makes up the ROI volume.

The voxels are grouped into a given number of groups, eg k different groups as shown by equation (14).

That is, the k groups of equation (14) are made up of the v¡ voxels closest to the centroid μ _i of the group. The centroids are computed from all j voxels in such a way that the sum of all the distances to the centroids is the minimum possible. This is expressed mathematically as shown in equation (15).

The previous model is solved as an optimization problem and, once the centroids have been established, each voxel will have a distance to the nearest centroid.

Then, in the stage of creating the matrix of characteristics of the structures of the nodules or tumors with the distance vector that is observed in Figure 4 with the number 6, where each one of the characteristics of each nodule will be attached to the matrix of characteristics that will make up the training matrix, with the values that characterize the tumor structures for the benefit of the census of the attributes of invasiveness in nodules and tumors, thus defining a distance vector that will be thresholded to the final risk scale.

The descriptors and the distance based on K-means make up a feature vector x for each voxel of the node, these descriptors being the input of an artificial neural network. The network is trained for the classification of each voxel, generating a new risk scale. The neural network would have the following output for the first layer with m artificial neurons:

Where x is the n-dimensional feature vector, that is, x ∈ R ⁿ . W ⁽¹⁾ is the matrix of weight parameters or training parameters, this being W ⁽¹⁾ ∈ R ^mxn . b ⁽¹⁾ is the bias vector of the neuronal model and f ⁽¹⁾ is the activation function.

In the stage of determining the biomarker of invasiveness that is observed in Figure 5 with number 7, it is found that the new matrix of characteristics by means of the deep learning algorithm makes it possible to predict the objective variable that will be the biomarker of tumor invasiveness, taking into account most of the biomarkers such as: class, type, diameter, geometry, map of pixel distances (tumor invasiveness), among others.

In the stage of determining the risk of the nodule that is observed in Figure 5 with the number 8, a probability distribution map of the values found by the distance map is made, obtaining a new risk scale.

Basically, the previous steps obtain a matrix of characteristics of the objective variable of the invasive biomarker, achieving a new AI or deep learning algorithm, with the ability to predict, taking into account most of the biomarkers such as: Class, type, diameter, geometry, map of pixel distances (tumor invasiveness). Accordingly, the objective variable will be the tumor invasiveness that will be projected with this new approach and will be ready to be thresholded to a new risk scale, to make a call to action from the system, correlating the Fleischner glossary.

Finally, the new risk scale is expressed by means of a probability distribution map of the values found by the distance map, where the new risk scale correlated with Fleischner, will be able to quickly communicate the invasiveness of a tumor, with respect to the transformation in the time of the intensity of the radio opacity in each of the non-solid nodular structures or in ground glass, as well as in the first presences of invasive factor, where it will be possible to map the first invasive findings in situ or the transformation to an invasive adenocarcinoma and its morphological conformations of intensities by degree of neighborhood.

In other words, the new risk scale is correlated with the Fleischner atlas, which will indicate the invasiveness of a tumor, with respect to the transformation over time of the intensity of the radio opacity in each of the non-solid nodular structures or in ground glass, as well as, in the first presences of invasive factor, where it will be possible to map the first invasive findings in situ or the transformation to an invasive adenocarcinoma and its conformations. morphological intensities by degree of neighborhood.

The fully connected artificial neural network follows the same behavior as the convolutional neural network. Therefore, for the following layers, the same model of equation (16) is followed, but with the difference that the input of each layer is the output of the previous layer, as shown in equation 7).

Consequently, for a fully connected network with p layers and q outputs, it would have the form expressed in equation (18).

Where yout ^∈ R ^q , each output being a probability value for each category of the new risk scale.

I I I . SI SYSTEM TO CHARACTERIZE LUNG TUMORS

Figure 6 shows the system (100) to characterize lung tumors (solid, subsolid and ground glass) based on invasive criteria using pixel distance and deep learning algorithms in CT scans of the present invention includes an imaging device (101) to form images of a target, for example, an axial tomograph computer, radiology equipment, mammographs, etc. ; an input interface ( 103) such as a screen, keyboard, mouse, etc. ; at least one processor ( 104) with a viewer with a vendor-unrelated (VNA) file or thoracic computed axial tomography reconstruction modality; a database (102) where the information that identifies each subject is stored, such as their medical history, chest studies, among others; a memory (105) coupled to at least one processor, an output interface (106), where the memory has executable instructions to implement in at least one processor the method comprising: i. Receive chest computed tomography images of subjects showing lung lesions, nodules or tumors or any type of abnormality;

To perform any of the functions described in this document, the processor (104) can execute one or more instructions, such as program modules where the deep learning algorithms that are applied in the development of the method of the invention are found, stored in one or more computer-readable storage media (for example, memory 105), which store instructions for execution by the processor (105).

Generally, program modules include routines, algorithms, objects, components, data structures, etc. that perform the particular tasks or implement particular data types of the present invention. While the systems and methods of the present disclosure have been particularly shown and described with reference to exemplary embodiments thereof, it will be understood by those skilled in the art that various changes in form and detail may be made without departing from the scope of the present disclosure.

EXAMPLE

It begins with the performance of a computerized axial tomography by means of a tomograph, to subjects that show lesions, nodules or lung tumors or any type of anomaly, storing the images and their clinical histories in the database.

The processor with the Indira® VNA viewer downloads the computed axial tomography images on the screen, such as Figure 2, which are stored in the database, and starts executing the instructions that are stored in memory to implement the method that comprises: i. Receive chest computed tomography images of subjects that show lesions, nodules or lung tumors or any type of anomaly;

In fact, the characterization of incidental nodes by the characterization of distances per group of pixels allows the final representation of the map of pixel distances per group applied to the non-small cell cancer segment, such as: adenocarcinoma structures, squamous cell cancer structures, as well as undifferentiated structures, to project a fine conformation of the tumor structure, differentiating the invasive factor from the rest of the structure in density to perform early risk thresholding, establishing management protocols in the face of imminent aggressiveness of a nodule or tumor.

The present invention has a wide application in bioengineering since it makes projections of physical-mathematical methods, where this development is synergistically articulated with computerized axial tomography modalities, as well as the modification of the subject prioritization processes involved in PACS-type systems, which are usually integrated into these tools.

Also, in RIS-PACS type systems, they are one of the technological artifacts that can exploit this technology the most, since it can act in the workflow of the characterization and description of the pathological event that is usually carried out in a RIS, as well as, it will be able to articulate its outputs to a PACs system, which will be able to manipulate the work list of radiologists, ensuring a list of priorities on an alert process, determined by the output parameter of the system through the invasiveness of a resulting tumor.

In medical image viewers, they are an excellent development niche, where this mechanism can be used as a second opinion process, increasing the certainty coefficient in the search for incidental findings in lung cancers, as well as finally describing the call to action through the risk threshold through the invasiveness of the tumor or nodule.

In addition, in advanced post-processing it can be used as a tool to enhance the characteristics of invasive pathological traits, which help to differentiate, through a distance map, those groups of pixels that have been discovered, through the system and serving as a report to a radiologist for diagnostic purposes.

Finally, in tumor research it is evident that distance maps are capable of mapping all those intermediate scales that are ignored, especially in non-small cell and squamous cell cancers, where a projection is necessary to help determine the degree of invasiveness and the prognosis of aggressiveness of the structure to be studied, evidencing new patterns to observe those entries that can give more information about the life expectancy of the final patient.

Claims

REI VI N DI CATI ON ES

1 . A system (100) to characterize lung tumors based on invasive criteria using pixel distance and deep learning algorithms comprising:

An imaging device (101) for imaging a target;

A database (102) where the information that identifies each subject is stored;

an input interface ( 103);

At least one processor (104) programmed to characterize lung tumors (solid, subsumed, and ground glass) based on invasive criteria using pixel distance and deep learning algorithms;

A memory (105) coupled to at least one processor;

an exit interface (106);

Wherein the characterization of lung tumors comprises receiving computed tomography images of the chest from subjects that show lesions, nodules or lung tumors or any type of anomaly; determine the volume of the tumor structure from the chest images of the region of interest; segmenting the voxel representation by means of a learning algorithm; determine the projected final volumetry; calculate the distance maps by means of a second unsupervised classification algorithm; create the matrix of characteristics of the structures of the nodules or tumors with the distance vector; determine the biomarker of invasiveness; and determine the risk of the nodule.

2. A system (100) to characterize lung tumors based on invasive criteria using pixel distance and deep learning algorithms according to claim 1, characterized in that the imaging device comprises a computational axial tomograph, radiology equipment, mammographs, etc.

3. A system (100) to characterize lung tumors based on invasive criteria using pixel distance and deep learning algorithms according to claim 1, characterized in that at least a processor comprises a viewer with a file not linked to any provider or modality of reconstruction of a thoracic computed axial tomography.

4. A method to characterize lung tumors based on invasive criteria using pixel distance and deep learning algorithms comprising: i. Receive chest images of subjects that show lesions, nodules or lung tumors or any type of anomaly;

5. A method for characterizing lung tumors based on invasive criteria using pixel distance and deep learning algorithms according to claim 4, characterized in that the chest images are computed tomography (CT), magnetic resonance imaging, mammography and/or ultrasound.

6. A method to characterize lung tumors based on invasive criteria using pixel distance and deep learning algorithms according to claim 4, characterized in that the determination of the volume of the tumor structure of the thorax images is carried out by means of the projection of 3D patches of the region of interest (ROI), which takes representation slices in voxels guaranteeing the generalized volume of the tumor.

7. A method for characterizing lung tumors based on invasive criteria by means of pixel distance and deep learning algorithms according to claim 4, characterized in that the segmentation of the voxel representation by means of a learning algorithm is carried out by means of an encoder-decoder type deep learning network algorithm, where the stacking of convolutional filters allows highlighting the characteristics of the nodule in relation to its morphology and contour.

8. A method for characterizing lung tumors, based on invasive criteria using pixel distance and deep learning algorithms according to claim 7, characterized in that the segmentation of the voxel representation by means of a deep learning algorithm applies an encoder-decoder type architecture such as the U-Net network, which allows preserving the pixel distribution through image geometries and morphologies, projecting a new spatial domain of matrix representation, belonging to the pix group. The elements that make up the segmentation in correlation to the pathological categories of nodules and tumors.

9. A method for characterizing lung tumors based on invasive criteria using pixel distance and deep learning algorithms according to claim 7, characterized in that the U-Net network consists of an encoder and a decoder, which are divided into substages, each substage is made up of a series of convolutional filters before changing the size of the feature maps through clustering in the encoder or transposed convolution in the decoder.

10. A method for characterizing lung tumors based on invasive criteria using pixel distance and deep learning algorithms according to claim 4, characterized by calculating the distance maps, is performed using a second unsupervised classification algorithm with the modification of some topological criteria with K-means or Boltzmann machines in which the distance maps are obtained by pairs of pixels that can project the quantitative values that will demarcate the new risk thresholds.

eleven . A method for characterizing lung tumors based on invasive criteria using pixel distance and deep learning algorithms according to claim 4, characterized in that creating the matrix of characteristics of the structures of the nodules or tumors with the distance vector is done using the descriptors and the distance based on K-means make up a vector x of characteristics for each voxel of the nodule, said descriptors being the input of an artificial neural network.

12. A method for characterizing lung tumors based on invasive criteria using pixel distance and deep learning algorithms according to claim 4, characterized in that the determination of the biomarker of tumor invasiveness is determined by means of the deep learning algorithm taking into account biomarkers of class, type, diameter, geometry, map of pixel distances (tumor invasiveness), among others.

13. A method to characterize lung tumors based on invasive criteria using pixel distance and deep learning algorithms according to claim 4, characterized in that the determination of the risk of the nodule is carried out with a probability distribution map of the values found by the distance map and the correlation with the Fleischner atlas.