LU505348B1

LU505348B1 - A deep learning-based method for CT pulmonary angiography cardiac chamber segmentation and quantification

Info

Publication number: LU505348B1
Application number: LU505348A
Authority: LU
Inventors: Haoyu Yang; Anqi Liu; Min Liu; Linfeng Xi; Yifei Ni; Jianping Wang
Original assignee: China Japan Friendship Hospital
Priority date: 2023-10-25
Filing date: 2023-10-25
Publication date: 2024-04-25

Abstract

The present invention pertains to the field of medical technology and discloses a deep learning-based method for CT pulmonary angiography cardiac chamber segmentation and quantification. The method includes the following steps: Coarse localization of the heart; Cardiac segmentation; Utilization of a U-shaped architecture known as "3D U-Net," which is highly favored in medical segmentation, for cardiac segmentation in our research; the 3D U-Net employs an encoder-decoder architecture. The encoder part is used for hierarchical feature extraction, while the decoder part is employed to merge features for precise cardiac structural segmentation; Evaluation of the performance of the segmentation model; Statistical analysis and software; Excellent performance in cardiac localization and segmentation based on the 3D U-Net model; Good concordance and consistency in volume calculation compared to manual segmentation, with segmentation based on deep learning; the 3D U-Net approach based on deep learning enables automatic segmentation of cardiac structures.

Description

DESCRIPTION LU505348

A DEEP LEARNING-BASED METHOD FOR CT PULMONARY

ANGIOGRAPHY CARDIAC CHAMBER SEGMENTATION AND

QUANTIFICATION

TECHNICAL FIELD

The present invention pertains to the field of medical technology and specifically involves a deep learning-based method for CT pulmonary angiography cardiac chamber segmentation and quantification.

BACKGROUND

The present invention pertains to the field of medical technology, specifically addressing a deep learning-based method for CT pulmonary angiography cardiac chamber segmentation and quantification. The size and volume of the heart chambers are influenced by various pathological and physiological conditions and can be indicative of cardiovascular health.

Computed Tomography Pulmonary Angiography (CTPA), without electrocardiogram gating, is a crucial method for diagnosing and assessing pulmonary vascular diseases, such as pulmonary embolism (PE) or pulmonary hypertension (PH). It not only detects thrombotic loads but also quantifies cardiovascular parameters, including the transverse diameter or area of the right/left ventricles and the diameter of the main pulmonary artery, to provide insights into the state of the heart. Right ventricular enlargement 1s a key indicator of right ventricular dysfunction. Currently, these parameters require manual measurements and depend on the choice of CTPA slices. In contrast, cardiac chamber volume 1s a more objective parameter for evaluating heart size and function, as it is not dependent on the experience of radiologists or the selection of CTPA slices. However, automatic acquisition of cardiac chamber volumes is not currently possible. Automatic and accurate segmentation and assessment of cardiac chamber volumes are essential for clinical evaluation of disease severity.

To address this issue, we propose a deep learning-based method for CT pulmonary angiography cardiac chamber segmentation and quantification. This method robustly segments the cardiac structures of pulmonary embolism patients without any manual adjustments, enabling early, accurate, and fast diagnosis of cardiac function. This approach LU505348 has been applied in clinical practice, enhancing diagnostic efficiency and improving patient care.

SUMMARY

The purpose of the present invention is to provide a deep learning-based method for CT pulmonary angiography cardiac chamber segmentation and quantification, addressing issues in the prior art.

To achieve the aforementioned purpose, the present invention provides the following technical solution: a deep learning-based method for CT pulmonary angiography cardiac chamber segmentation and quantification, comprising the following steps:

S1: Coarse localization of the heart. S2: Cardiac segmentation. S3: Cardiac structural segmentation. S4: Evaluation of the performance of the segmentation model. S5: Statistical analysis and software.

In a preferred embodiment, in Step S1, considering the anatomical relationship between the heart and the lung field, we obtain the approximate range of the heart, completing the rough extraction of the cardiac region.

In a preferred embodiment, in Step S2, given a patient's CTPA slice sequence, we first extract a three-dimensional bounding box from the image based on anatomical priors.

Subsequently, the bounding box undergoes preprocessing and data processing. When this processed bounding box is provided as input to the model trained, it outputs the probability of the foreground or the heart for each voxel using a sigmoid function. Voxels with probabilities exceeding a threshold are assigned to the heart, resulting in the final segmentation produced by the model.

In a preferred embodiment, in Step S3, the inputs, segmentation results, and preprocessing labels in the cardiac segmentation stage are initially processed by extracting a bounding box that encompasses the entire manually segmented heart plus an additional 5 voxels of symmetric margin. Then, all extracted volumes are scaled to a size of 160x160x160 voxels. During the training phase, the data used for cardiac segmentation is also utilized for model training and model selection in the cardiac structural segmentation stage.

In a preferred embodiment, in Step S4, to evaluate the automatic segmentation, the LU505348 obtained cardiac region and cardiac structures are compared with manual ground truth annotations.

In a preferred embodiment, in Step S5, the interclass correlation coefficient (ICC) based on a single evaluator, absolute agreement, and a two-way random-effects model is used to assess the accuracy and segmentation performance of cardiac structural volume segmentation.

In a preferred embodiment, after model training in Step S3 is completed, a trained model is selected for application to two test cohorts. Further details about model training and selection are described in the supplementary materials. During testing, given a CTPA volume from a patient, the process begins with coarse cardiac localization and cardiac segmentation to obtain a cardiac prediction result. This is achieved by extracting a bounding box encompassing the entire prediction plus an additional 5 voxels of symmetric margin and scaling it to a size of 160x160x160 voxels. Preprocessing is applied to the model input and the cardiac segmentation mask from the previous step.

In a preferred embodiment, in Step S2, before training, images and labels obtained through coarse cardiac localization are preprocessed by resampling to a 1 mm isotropic voxel size and scaling to a 160x224x224 matrix.

The provided deep learning-based method for CT pulmonary angiography cardiac chamber segmentation and quantification offers the following significant advantages: 1. The method, based on the 3D U-Net model, excels in cardiac localization and segmentation performance. 2. It demonstrates good concordance and consistency when comparing volumes calculated by automatic segmentation using deep learning to manual segmentation. 3. The method, employing the 3D U-Net model based on deep learning, enables automatic segmentation of cardiac structures, providing a valuable tool for clinical applications.

BRIEF DESCRIPTION OF THE FIGURES LU505348

Fig. 11s a flow chart of the present invention

Fig. 2 is a performance diagram of the internal and external test models of the present invention;

Fig. 3 is a detailed view of the 3D U-Net of the present invention.

DETAILED DESCRIPTION OF THE INVENTION

In order to provide a clearer understanding of the technical features, objectives, and effects of the present invention, the specific embodiments of the invention are described in conjunction with the accompanying drawings.

As shown in Figures 1-3, the deep learning-based method for CT pulmonary angiography cardiac chamber segmentation and quantification, as provided by the present invention, includes the following steps:

S1: Coarse localization of the heart; Taking into consideration the anatomical relationship between the heart and the lung field, we obtain the approximate range of the heart, completing the rough extraction of the cardiac region.

S2: Cardiac segmentation; A U-shaped architecture known as "3D U-Net" is widely popular in medical segmentation and is used in our research for cardiac segmentation. The 3D

U-Net employs an encoder-decoder architecture, where the encoder part (left portion) is used for hierarchical feature extraction, and the decoder part (right portion) is used to merge features to achieve accurate results. The structural details of this network are shown in Figure 2(B). For a given patient's CTPA slice sequence, we first extract a three-dimensional bounding box from the image based on anatomical priors. Subsequently, the bounding box undergoes preprocessing and data processing. When this processed bounding box is provided as input to the model we trained, it outputs the probability of the foreground or the heart for each voxel using a sigmoid function. Voxels with probabilities exceeding a threshold are assigned to the heart, resulting in the final segmentation produced by the model. Before training, images and labels obtained through coarse cardiac localization are preprocessed by resampling to a 1 mm isotropic voxel size and scaling to a 160x224x224 matrix.

S3: Cardiac structural segmentation; In the cardiac segmentation stage, the inputs, segmentation results, and preprocessing labels are initially processed by extracting a bounding box that encompasses the entire manually segmented heart plus an additional 5 voxels of LU505348 symmetric margin. Then, all extracted volumes are scaled to a size of 160x160x160 voxels.

During the training phase, the data used for cardiac segmentation is also utilized for model training and model selection in the cardiac structural segmentation stage. After model training, a trained model is selected for application to two test cohorts. Further details about model training and selection are described in the supplementary materials. During testing, given a

CTPA volume from a patient, the process begins with coarse cardiac localization and cardiac segmentation to obtain a cardiac prediction result. This is achieved by extracting a bounding box encompassing the entire prediction plus an additional 5 voxels of symmetric margin and scaling it to a size of 160x160x160 voxels. Preprocessing is applied to the model input and the cardiac segmentation mask from the previous step.

S4: Evaluation of the segmentation model's performance; To assess the automatic segmentation, the obtained cardiac regions and cardiac structures are compared with manual ground truth annotations.

SS: Statistical analysis and software; The interclass correlation coefficient (ICC) based on a single evaluator, absolute agreement, and a two-way random-effects model is used to evaluate the accuracy and segmentation performance of cardiac structural volume segmentation.

Results: Population Characteristics: A total of 558 patients were included in the study, of which 314 patients (age 22-87 years, average age 61.50 years) had CTPA scans used for model training, 150 patients (age 21-91 years, average age 63.07 years) had CTPA scans used for internal testing. Ninety-four patients from the Second Medical Center (age 19-87 years, average age 65.40 years) were used for external testing. Table 1 displays the clinical and demographic characteristics of the training and testing cohorts. The age and gender were comparable between the training dataset and the internal testing dataset (p > 0.05). There was an age difference between the training dataset and the external testing dataset (p < 0.05), while the gender was similar between the two groups (p > 0.05).

Table 1 Participant Demographics

Test queue

Variable Cultivate supporters Internal Test external Test supporter supporter

Total number 314 150 94

Age, age standard deviation 61.50 13.94 63.07 15.60 65.40 12.93

Age range, years 22-87 21-91 19-87

Race Chinese Chinese Chinese

Sex

Male 146 57 46

Female 168 93 48

Performance of segmented models in internal test queue:

Table 2 shows the evaluation results of the proposed method on an internal test cohort.

According to these results, DL-based automatic segmentation has good correlation with manual segmentation. Figure 1 shows that in the internal test cohort, the DSC of the trained model for cardiac segmentation was 0.940 (95% CI [0.938, 0.942]); the DSC of the model for cardiac structure segmentation at MLV was 0.900 (95% CI [0.894, 0.906]), 0.928 (95% CI [0.924, 0.931]) in LABC, 0.927 (95% CI [0.922, 0.932]) in LVBC, and 0.902 (95% CI [0.895, 0.908] in RABC ), 0.936 in the United States (936) Figure 4 shows an example of successful segmentation in the internal test cohort.

Table 2 Performance of internal tests

Heart structure segmentation

Mean LAB LVB RABC RVB American Seducing and lung C C C Statistical brainwashing

Heart volume Associatio the opposite

Metric segment n; sex and ation American chatting up

Standards artists (

Institute Pick-up Artist )

Dice 0.940 0.900 0.928 0.927 0.902 0.936 0.888 0.925 0.024 0.014 0.037 0.022 0.031 0.039 0.019 0.040

Remem 0.934 0.892 0.917 0.921 0.900 0.925 0.826 0.896 0.045 ber 0.022 0.043 0.035 0.043 0.052 0.036 0.074

Accurat 0.947 0.909 0.940 0.935 0.908 0.949 0.967 0.958 0.020 e 0.020 0.046 0.037 0.041 0.056 0.024 0.032

Note: All results are presented as means and standard deviations. MLV = Left

Ventricular Myocardium, LABC = Left Atrium Blood Chamber, LVBC = Left Ventricular

Blood Chamber, RABC = Right Atrium Blood Chamber, RVBC = Right Ventricular Blood

Chamber, ASA = Ascending Aorta, PUA = Pulmonary Artery Performance of the segmentation model in the external testing cohort: The evaluation results of our automated method are presented in Table 3. These results, along with Figure 1, demonstrate that in the external testing group, the proposed automatic segmentation remains highly correlated with the manual segmentation by radiologists. The Dice Similarity Coefficient (DSC) for the model used for cardiac segmentation is 0.936 (95% CI [0.933, 0.939]). For the model used in cardiac structural segmentation, the DSC for MLV is 0.892 (95% CI [0.881, 0.9031), for

LABC is 0.920 (95% CI [0.914, 0.9271), for LVBC is 0.904 (95% CI [0.892, 0.916]), and for

RABC is 0.895 (95% CI [0.888, 0.901]). Figure 5 illustrates successful segmentation in the LU505348 external testing group.

Table 3 Performance of external tests

Heart structure segmentation

Mean LAB LVB RABC RVB American Seducing lung C C C Statistical and volum Associati brainwashi

Heart e on; ng the

Metric segmentatio American opposite n Standards sex and

Institute chatting up artists (

Pick-up

Artist)

Dice 0.936 0.014 0.892 0920 0904 0.895 0.924 0.869 0.926 0.054 0.031 0.058 0.032 0.017 0.045 0.017

Reme 09220025 0880 0905 0905 0873 0.899 0.790 0.909 mber 0.062 0.045 0.072 0.040 0.036 0.077 0.029

Accur 09510.013 0907 0938 0910 0.920 0.951 0.972 0.946 ate 0.055 0.039 0.067 0.046 0.021 0.016 0.021

THIS

We have proposed a DL-based framework for whole-heart segmentation and cardiac structure segmentation using CTPA images. For cardiac structure segmentation, we referred to the results of the multi-modality whole-heart segmentation challenge and segmented seven cardiac structures, including MLV, LABC, LVBC, RABC, RVBC, ASA, and PUA, and automatically calculated their volumes. The experimental results validate the accuracy and robustness of this framework and yield two key findings:

The performance of cardiac localization and segmentation using the 3D U-Net model is excellent.

DL-based automatic segmentation demonstrates good concordance and consistency in LU505348 volume measurement compared to manual segmentation.

Cardiac localization forms the foundation for measuring the various cardiac chambers.

Initially, we use the 3D U-Net to segment the entire heart. In this stage, precise cardiac localization is performed rather than structural segmentation. The final selected cardiac segmentation model achieved an average DSC of 0.940 (95% CI [0.938, 0.942]) in the internal testing cohort and an average DSC of 0.936 (95% CI [0.933, 0.939]) in the external testing cohort. This indicates a high correlation between automatic and manual segmentation of the entire heart.

Next, we use the 3D U-Net to segment cardiac structures. The DSC ranges from 0.888 to 0.936 in the internal testing cohort and from 0.869 to 0.926 in the external testing cohort.

Previous research on semantic segmentation of pulmonary arteries achieved DSCs of 0.85 and 0.92, and this study obtained a DSC of 0.925 in the internal testing cohort, which is similar to these results.

In the external testing cohort, there was one case of partial failure in cardiac structure segmentation, with a DSC of 0.46 for MLV and 0.56 for LVBC, indicating poor segmentation performance. Insufficient contrast agent filling is the reason for segmentation failure.

Therefore, if DL models are to be applied in clinical practice, higher quality control for CTPA images is required. Other structures showed DSCs ranging from 0.77 to 0.94, indicating good segmentation performance. Despite one instance of segmentation failure, other cardiac structures exhibited high DSCs and good segmentation performance, suggesting a strong correlation between DL-based automatic segmentation and manual segmentation by physicians.

Volume measurements offer advantages over one-dimensional or two-dimensional measurements and apply to other body areas. The volume of each ventricle is an objective indicator for assessing cardiac function, determining pulmonary hypertension, and stratifying pulmonary embolism risk. However, because ventricular volume measurements on non-gated

CTPA images are not widely used in clinical practice, we used a DL-based method to automatically calculate these volumes.

Quantitative analysis of cardiac structure segmentation showed excellent performance, LU505348 with ICCs exceeding 0.9 for all structures except ASA (0.834 in the internal testing cohort and 0.707 in the external testing cohort). This result is consistent with the DSC results, indicating that both are good metrics for evaluating cardiac structure segmentation.

Additionally, sensitivity and precision were used to evaluate segmentation performance, and the results were similar to the DSC, indicating their utility for assessing cardiac structure segmentation.

Claims

CLAIMS LU505348

1. A deep learning-based method for CT pulmonary angiography cardiac chamber segmentation and quantification, characterized by the following steps: step 1 (S1): coarse localization of the heart step 2 (S2): cardiac segmentation step 3 (S3): cardiac structural segmentation step 4 (S4): evaluation of the performance of the segmentation model step 5 (SS): statistical analysis and software.

2. The method for CT pulmonary angiography cardiac chamber segmentation and quantification based on deep learning according to claim 1, characterized by the following: in step 1 (S1), considering the anatomical relationship between the heart and the lung field, the approximate range of the heart is obtained, achieving rough extraction of the cardiac region.

3. The method for CT pulmonary angiography cardiac chamber segmentation and quantification based on deep learning according to claim 1, characterized by the following: in step 2 (S2), given a patient's CTPA slice sequence, a three-dimensional bounding box is extracted from the image based on anatomical priors; the bounding box undergoes preprocessing and data processing; the trained model, when provided with this processed bounding box as input, outputs the probability of the foreground or the heart for each voxel using a sigmoid function; voxels with probabilities exceeding a threshold are assigned to the heart, resulting in the final segmentation produced by the model.

4. The method for CT pulmonary angiography cardiac chamber segmentation and quantification based on deep learning according to claim 1, characterized by the following: in step 3 (S3), the input, segmentation results, and preprocessing labels in the cardiac segmentation stage are initially processed by extracting a bounding box that encompasses the entire manually segmented heart plus an additional 5 voxels symmetric edge; subsequently, all extracted volumes are scaled to the size of 160x160x160 voxels; during the training phase, the data used for cardiac segmentation is also utilized for model training and model selection in the cardiac structural segmentation stage.

5. The method for CT pulmonary angiography cardiac chamber segmentation and LUS05348 quantification based on deep learning according to claim 1, characterized by the following: in step 4 (S4), to evaluate the automatic segmentation, the obtained cardiac region and cardiac structures are compared with manual ground truth annotations.

6. The method for CT pulmonary angiography cardiac chamber segmentation and quantification based on deep learning according to claim 1, characterized by the following: in step 5 (SS), the interclass correlation coefficient (ICC) based on a single evaluator, absolute agreement, and a two-way random-effects model is used to assess the accuracy and segmentation performance of cardiac structural volume segmentation.

7. The method for CT pulmonary angiography cardiac chamber segmentation and quantification based on deep learning according to claim 1, characterized by the following: in step 3 (S3), after model training is complete, a trained model is selected for application to two test cohorts; more details about model training and selection are described in the supplementary materials; during testing, given a CTPA volume from a patient, the process begins with coarse cardiac localization and cardiac segmentation to obtain a cardiac prediction result; this is achieved by extracting a bounding box encompassing the entire prediction plus an additional 5 voxels of symmetric margin and scaling it to a size of 160x160x160 voxels; preprocessing is applied to the model input and the cardiac segmentation mask from the previous step.

8. The method for CT pulmonary angiography cardiac chamber segmentation and quantification based on deep learning according to claim 1, characterized by the following: in step 2 (S2), before training, images and labels obtained through coarse cardiac localization are preprocessed by resampling to a 1 mm isotropic voxel size and scaling to a 160x224x224 matrix.