A deep learning-based ultrasound myocardial perfusion imaging diagnosis method

Abstract

The application discloses a deep learning-based ultrasonic myocardial perfusion imaging diagnosis method, which comprises the following steps: setting an image label, image preprocessing, image grouping, model training and result testing. The deep learning-based ultrasonic myocardial perfusion imaging diagnosis method has higher sensitivity and specificity and can have higher diagnostic value for myocardial perfusion imaging scoring.

Description

Technical Field

[0001] This invention relates to myocardial perfusion imaging technology, and more specifically to a deep learning-based ultrasound myocardial perfusion imaging diagnostic method. Background Technology

[0002] Currently, conventional two-dimensional echocardiography is a routine examination method in the diagnosis of coronary artery disease. However, this method is highly subjective and its diagnostic effectiveness is not ideal for ischemic cardiomyopathy without segmental wall motion abnormalities, or for ischemia with non-obstructive coronary artery disease (INOCA). Peripherally transvenous myocardial contrast ecocardiography (MCE) is a new technique for diagnosing myocardial perfusion at the microcirculation level. In recent years, it has been gradually promoted and applied in clinical practice, providing an effective way to improve the detection rate of myocardial ischemia and evaluate myocardial microcirculation. However, MCE diagnosis is highly specialized, has a high learning time cost, and cannot be widely implemented in all hospitals.

[0003] In recent years, computer-aided diagnostic (CAD) systems have achieved remarkable development. From classic machine learning (ML) methods to the currently popular deep learning (DL) models, the performance of CAD systems has greatly improved. Deep learning (DL), also known as deep neural learning networks, simulates human neural pathways, using multi-layered neural networks. Computer systems extract large amounts of more complex and hidden data features and simultaneously train classifiers. These artificial neural networks can be hundreds of layers deep, where data is processed continuously until the final output is obtained. Due to their large model capacity and deeper, more complex structure, they have better accuracy and wider applicability, overcoming the subjectivity of human ultrasound physicians and demonstrating diagnostic efficacy comparable to that of human ultrasound physicians in multiple studies. Summary of the Invention

[0004] The main objective of this invention is to provide a deep learning-based ultrasound myocardial perfusion imaging diagnostic method with higher sensitivity and specificity, and higher diagnostic value for myocardial perfusion imaging scores.

[0005] The technical solution adopted in this invention is: a deep learning-based ultrasound myocardial perfusion imaging diagnostic method, comprising: Set image labels: Set the label of an image with an infusion score of 1 to 0, the label of an image with a score of 2 to 1, and the label of an image with a score of 3 to 2; Image preprocessing: The original dataset images were 1024×768 pixels in size. The images were resized to 256×256 pixels for image enhancement. The original myocardial perfusion imaging images consisted of 860 images. Data enhancement was performed using rotation. Three additional samples were generated from each original image using rotation angles of 90 degrees, 180 degrees, and 270 degrees, increasing the original dataset from 860 images to 3440 images. Image grouping: 230 images are randomly selected from each of the three types of images, for a total of 690 images as the test set, which accounts for about 20% of the total number of images. The remaining 80%, 2750 images, are used as the training set, ensuring that images in the training set do not appear in the test set. Model training: The training set images are trained and tested on different networks to obtain the loss and accuracy on the training set and validation set. The model parameters with the highest accuracy during training are selected. The number of classes 0 and 2 is much smaller than that of class 1, which will affect the training results. Therefore, the dataset is weighted during each training session to make the ratio of the number of images of the three classes close to 1:1:1. Results testing: The model parameters with the highest accuracy selected during model training are used to test the test set to obtain the final test results.

[0006] Furthermore, the evaluation index formula for the deep learning-based ultrasound myocardial perfusion imaging diagnostic method is shown in the following equation:

[0007]

[0008]

[0009] Where TP refers to the number of positive samples that are correctly classified; FP refers to the number of negative samples that are misclassified as positive samples; TN refers to the number of negative samples that are correctly classified as negative samples; and FN refers to the number of positive samples that are misclassified as negative samples.

[0010] Advantages of this invention: In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description

[0011] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0012] Figure 1 This is the image classification standard diagram of the present invention; Figure 2This is the ROC curve of the diagnostic accuracy of the present invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0014] Materials and Methods: Research subjects: From January 2022 to June 2023, 263 patients underwent myocardial perfusion imaging at Shaanxi Provincial People's Hospital. Among them, 171 were male and 92 were female, with a mean age of (54.34±10.34) years. Inclusion criteria: (1) no contraindications for myocardial perfusion imaging; (2) clear images without significant gas interference, which can be further analyzed.

[0015] Ultrasound examination equipment: This study used the Philips EPIQ CVX type ultrasound diagnostic instrument and the S5-1 adult cardiac probe.

[0016] Myocardial perfusion imaging image acquisition and database establishment: To ensure the quality of the images included in the study, all ultrasound images were acquired by ultrasound physicians with more than 5 years of clinical experience.

[0017] An intravenous line was established for the patient, and an electrocardiogram (ECG) was connected for MCE examination. Diluted perfluoropropane human serum albumin microspheres (Lipin Pharmaceutical) contrast agent was slowly injected via the patient's left median cubital vein. Images were acquired in the apical four-chamber, two-chamber, and three-chamber views, as well as the long-axis and short-axis views of the left ventricle. At the start of dynamic image acquisition, the contrast agent microbubbles in the myocardial tissue were burst using the FLASH function to record the myocardial refill process. Thirteen cardiac cycles were acquired for each section. After dynamic image acquisition, the myocardial perfusion images of each section were played back frame by frame. Static myocardial perfusion images of the left ventricle during diastole were also acquired. All static images were uploaded to the image archiving and communication system. The final dataset included 860 images of the left ventricular four-chamber, apical three-chamber, apical two-chamber, and left ventricular long-axis views, each image measuring 1024*768 pixels.

[0018] 1.2.3 Routine ultrasound diagnosis: Three ultrasound physicians with more than 15 years of clinical experience independently reviewed the images in a blinded and randomized manner, marked the myocardial perfusion score, and conducted a second joint review of the images for any inconsistent judgments, reached a consensus, and used it as the standard for determining the myocardial perfusion score.

[0019] Myocardial perfusion levels were scored according to the American Society of Echocardiography's 17-segment left ventricular division: complete perfusion was negative, scoring 1 point; compared to a completely perfused segment, the segment showed incomplete contrast enhancement, scoring 2 points; compared to a completely perfused segment, the segment showed no contrast enhancement, scoring 3 points (see Table 1). Figure 1 ).

[0020] Table 1. Introduction to Scoring Criteria

[0021] A deep learning-based ultrasound myocardial perfusion imaging diagnostic method includes: Set image labels: Set the label of an image with an infusion score of 1 to 0, the label of an image with a score of 2 to 1, and the label of an image with a score of 3 to 2; Image preprocessing: The original dataset images were 1024×768 pixels in size. The images were resized to 256×256 pixels for image enhancement. The original myocardial perfusion imaging images consisted of 860 images. Data enhancement was performed using rotation. Three additional samples were generated from each original image using rotation angles of 90 degrees, 180 degrees, and 270 degrees, increasing the original dataset from 860 images to 3440 images. Image grouping: 230 images are randomly selected from each of the three types of images, for a total of 690 images as the test set, which accounts for about 20% of the total number of images. The remaining 80%, 2750 images, are used as the training set, ensuring that images in the training set do not appear in the test set. Model training: The training set images are trained and tested on different networks to obtain the loss and accuracy on the training and validation sets, and the model parameters with the highest accuracy during training are selected. Considering the imbalance of data classes, the number of classes 0 and 2 is much less than that of class 1, which will affect the training results. Therefore, the dataset is weighted during each training session to make the ratio of the number of images of the three classes close to 1:1:1. Results testing: The model parameters with the highest accuracy selected during model training are used to test the test set to obtain the final test results.

[0022] The evaluation index formula for the deep learning-based ultrasound myocardial perfusion imaging diagnostic method is shown in the following equation:

[0023]

[0024]

[0025] Where TP refers to the number of positive samples that are correctly classified; FP refers to the number of negative samples that are misclassified as positive samples; TN refers to the number of negative samples that are correctly classified as negative samples; and FN refers to the number of positive samples that are misclassified as negative samples.

[0026] Comparison of the diagnostic efficacy of deep learning models with ultrasound physicians of different experience levels: All test data were reviewed by three groups of clinical ultrasound physicians. The three groups of ultrasound physicians were divided into junior (<5 years), middle-aged (5-10 years), and senior (>10 years) groups according to their work experience. They made independent diagnoses in a blinded and randomized manner, scored myocardial perfusion and labeled the type, and recorded the review time.

[0027] Statistical methods: Data analysis was performed using SPSS 25.0 software. Quantitative data were expressed as mean ± standard deviation (x±s), and categorical data were expressed as frequency or percentage. The chi-square test was used for comparison. The Delong Test was used to calculate the receiver operating characteristic (ROC) curve to calculate diagnostic efficacy. A p-value < 0.05 was considered statistically significant.

[0028] result: Study population: This study included 263 patients and 3440 images. The images were divided into a training set and a test set. The test set contained 690 images, and the training set contained 2750 images.

[0029] Performance of deep neural network models on the test set: The deep learning model demonstrated good performance in myocardial perfusion scoring, achieving an overall accuracy of 77.536%. The sensitivities for classes 0, 1, and 2 were 77.391%, 72.609%, and 82.609%, respectively, while the specificities were 89.783%, 82.174%, and 94.348%, respectively. (Table 3) Table 3: Diagnostic efficacy of deep learning models

[0030] A comparison of the diagnostic efficacy of deep learning models and ultrasound physicians: Table 4 shows a comparison of the diagnostic efficacy of the deep learning model and the ultrasound physicians in each group. Figure 2 .

[0031] Table 4: Performance of ultrasound physicians with <5 years, 5-10 years, and >10 years of clinical experience in diagnosing myocardial perfusion abnormalities on the test set.

[0032] *: p-value < 0.05 compared to the model group; **: p-value > 0.05 compared to the model group.

[0033] ROC curve for diagnostic accuracy is as follows: Figure 2 As shown.

[0034] Figure 2 In China: the area under the curve (AUC) for the model group was 0.836, for the junior physician group it was 0.815, for the mid-level physician group it was 0.860, and for the senior physician group it was 0.883.

[0035] The sensitivity of the degree-learning model in determining myocardial perfusion score reached 89.783%, comparable to that of intermediate-level ultrasound physicians (89.783% vs 88.913%, P > 0.05), superior to that of junior-level ultrasound physicians (89.783% vs 83.913%, P < 0.05), but lower than that of senior-level ultrasound physicians (89.783% vs 93.913%, P < 0.05). The AUC was also comparable to that of intermediate-level ultrasound physicians and higher than that of junior-level ultrasound physicians. However, the specificity showed no statistically significant difference compared to the three groups of ultrasound physicians (all P > 0.05).

[0036] Furthermore, the average diagnostic time for the deep learning model was (1.47±0.75) s per case. The mean time for the three groups of ultrasound physicians to judge each image was (15.21±7.92) s, (11.53±5.36) s, and (8.46±4.59) s, respectively, with an average of (11.73±6.39) s. The deep learning model reduced the judgment time per image by 10.26 s, improving diagnostic efficiency by 7 times.

[0037] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based ultrasound myocardial perfusion imaging diagnostic method, characterized in that, include: Set image labels: Set the label of an image with an infusion score of 1 to 0, the label of an image with a score of 2 to 1, and the label of an image with a score of 3 to 2; Image preprocessing: The original dataset images were 1024×768 pixels in size. The images were resized to 256×256 pixels for image enhancement. The original myocardial perfusion imaging images consisted of 860 images. Data enhancement was performed using rotation. Three additional samples were generated from each original image using rotation angles of 90 degrees, 180 degrees, and 270 degrees. The original dataset was increased from 860 images to 3440 images. Image grouping: 230 images are randomly selected from each of the three types of images, for a total of 690 images as the test set, which accounts for about 20% of the total number of images. The remaining 80%, 2750 images, are used as the training set, ensuring that images in the training set do not appear in the test set. Model training: The training set images are trained and tested on different networks to obtain the loss and accuracy on the training and validation sets, and the model parameters with the highest accuracy during training are selected. Since the number of classes 0 and 2 is much smaller than that of class 1, it will affect the training results. Therefore, the dataset is weighted during each training session to make the ratio of the number of images of the three classes close to 1:1:

1. Results testing: The model parameters with the highest accuracy selected during model training are used to test the test set to obtain the final test results.

2. The deep learning-based ultrasound myocardial perfusion imaging diagnostic model according to claim 1, characterized in that, The evaluation index formula for the deep learning-based ultrasound myocardial perfusion imaging diagnostic method is shown in the following equation: ， ， ， Where TP refers to the number of positive samples that are correctly classified; FP refers to the number of negative samples that are misclassified as positive samples; TN refers to the number of negative samples that are correctly classified as negative samples; and FN refers to the number of positive samples that are misclassified as negative samples.

Images