We propose a classification convolutional neural network for performing CAT detection from CAG. This code is based on the image classification module available in the DEEP Open Catalog.13. The original classification model developed in the DEEP framework was adapted for our specific task.
Data collection
This is a retrospective study approved by the Human Research Ethics Committee of Maule Health Authority and the Biomedical Research Ethics Committee of Talca Regional Hospital, Chile. All methods reported in this study were performed in accordance with relevant guidelines and regulations. As this study was approved by the Biomedical Research Ethics Committee of Talca Regional Hospital, there was no direct patient interaction and no informed consent was required.
population
This is a retrospective clinical study of a sample of 18,000 patients who were referred to the hemodynamics department of Talca Regional Hospital between 2016 and 2022 with symptoms of coronary artery disease. Subjects underwent CAG with a diagnosis that did not reveal significant angiographic lesions. That is, the patient reported no coronary artery changes related to the clinical condition for which testing was required. The patient population was from Chile, with a mean age of 68 years (SD 8 years), consisting of 216 males with a mean age of 69 years (SD 9 years) and 185 females with a mean age of 68 years (SD 6 years). I was. ). Of these patients, 658 his CAGs were considered according to the inclusion and exclusion criteria as described in the ‘Initial inclusion and exclusion criteria’ section. A participating patient’s CAG was obtained anonymously. Medical records were obtained from the database of Talka Regional Hospital. Comorbidities were not considered in patient selection.
Initial inclusion and exclusion criteria
Initial inclusion criteria were: Subjects without angiographically disfiguring lesions or significant anatomic changes, left 45°/25° (spider), right 45°/A 0° projection with an angular variance of 3° or less. Exclusion criteria were patients with left ventricular hypertrophy, valvular heart disease, coronary artery anatomic variations, dysmorphic coronary anomalies, previous history of CAG, cardiomyopathy, or history of other cardiac disease, and these patients This is because they may have pathologically abnormal coronary arteries. Most of these criteria were proposed by14. After applying all inclusion and exclusion criteria, the final sample consisted of 658 CAG images.
Calibration, patient and image selection
The image collections corresponding to the selected patients were obtained from databases of two different angiographers affiliated with the Department of Hemodynamics at Talka Regional Hospital. Images from 2016 to 2019 were obtained from Siemens® Angiographer (95 CAG), 2019-2022 images from Philips® Angiographer (563 CAG).Only images corresponding to left 45°/caudal projection 25° (Spider) and left 45° of left coronary artery/0° of right coronary artery projection was chosen.
One image capture was obtained from each angiographic film. Files were saved in jpg or png format for left and right her CAG at the time of maximal arterial contrast loading. Next, the images were de-artifacted where necessary, as foreign objects interfere with the interpretation of the neural network model. Coronary artery tortuosity (CAT), identified by the presence of 3 or more consecutive tortuosities (defined as a 45-degree change in vessel orientation) along at least one major epicardial coronary artery trunk, is not visible on imaging. It was thought to be labeled as: Equivalent to CAT patients.
Fifty angiographic images from a total set meeting pre-established inclusion and exclusion criteria were used for calibration. As a baseline, results obtained by 3 specialists in the field (3 cardiologists and vascular specialists) were considered and reviewed by cardiologists using the following qualitative methods: (a a) visual inspection of vessel tortuosity by defining fixed anatomical points within the vessel which torsion to measure; and (b) recording the number of inflection points between fixed anatomical points. , (c) counting the number of torsions and loops and classifying them according to defined torsion criteria, (d) relating torsion to arterial elongation, and wall weakening observed as subtle changes in wall contrast uptake. will be The biometric analysis confidence for this initial calibration set was 98..0% for between-appraiser and within-appraiser values using intraclass correlation (ICC).
ground truth
Three cardiologists and angiologists who had been trained on the 50 angiographic images used for calibration used CAG was assessed to find CAT that functioned and among patients during the study period. Ground Truth was independently tested by her one of three cardiologists and vascular specialists with an average of 15 years of experience (her 12, 16, and 17, respectively) and an average of 7,000 examinations. established.
data set
Our final experimental dataset consisted of 658 CAG images corresponding to a total of 401 different patients. Table 1 shows the number of patients available for each type of CAG.
model design
A convolutional neural network (CNN) is a type of DL neural network specifically designed for analyzing images using both numerical (regression) or categorical (classification) labels. In this study, we had to solve a binary classification problem on CAG images to detect patients with either coronary tortuosity (CAT) or no coronary tortuosity (WCAT).
There are three sets (Training, Validation, and Testing) into which the data (images in this case) are subdivided for use in the DL model. A training set is used to train the model. During the training phase, hyperparameters are tuned to optimize model performance across the validation set. Training stops when the accuracy on the validation set no longer increases for a given number of epochs in the training phase. This is one of the most common normalization techniques in DL and is known as early stopping.15. The test set contains images that the model has never seen before and is used to assess the final unbiased accuracy of the model.
The procedure to detect CAT consisted first of training a CNN model using the CAG images. We trained 5 different models according to 5 cross-validation strategies16. The total number of patients was balanced in each of the five sets (training, validation, testing). This means that half of the images correspond to patients with coronary tortuosity (CAT) and the other half to patients without coronary tortuosity (WCAT) (Spider or 45). Images were randomly selected from one of the angiographers and consisted of 450 for training (225 CAT, 225 WCAT), 46 for validation (23 CAT, 23 WCAT), and 48 for testing (23 CAT, 23 WCAT). 24 CAT and 24 WCAT) with the same distribution for each cross-validation fold. Since the number of WCAT images was higher than the number of his CAT images, we randomly repeated 38 corresponding his CAT images in training in order to use all available her WCAT images. Both coronary artery projections are contained in the same CNN. It ran several tests to assess whether the DL model performed better when trained with each projection separately and when trained using both, and can show similar validation metrics. because it turned outBoth decided to train at 45° left/25° (spider), right 45°/The 0° projection includes more images during training, so the model can learn the difference between CAT and WCAT more accurately.
I used Xception17 A CNN with images of size 528 by 528 pixels. In our method, the CAG images were resized to meet this requirement. After experimenting with different model initializations, the batch size was finally set to 16 and the number of epochs was fixed at a maximum of 50, but early stopping was used to prevent overfitting (15 epochs per patient The set). As a result, the number of epochs was fixed at a maximum of 50. The percentage of training epochs was lower for the model. Also adopted the Adam optimizer18,19 to speed up training. First, a pretrained ImageNet base model was loaded to optimize the learning task.This methodology is known as transfer learning20. We used ImageNet’s pre-trained model and replaced the final fully connected (FC) layer with an FC layer that was adapted to the problem (binary classification). We then trained everything end-to-end, but the basic feature extractor trained at a much lower learning rate compared to the FC layer started from random weights. Fine-tuning the feature extraction function increases the relevance of features present in this particular problem that may not be visible on ImageNet.twenty one.
The deep learning architecture developed in this study to detect CAT is shown in Figure 1. We used data augmentation because the total number of images available was very limited.twenty two We improve the model on both the training and validation sets (see Table A1 for details).
Proposal of CNN architecture for coronary tortuosity detection.
Each model was trained using a Tesla V100-PCIE-32 GB GPU.The model was coded using Kerastwenty three and TensorFlow version 1.14.0twenty four Ubuntu 18.04.2 LTS.
statistical analysis
A five-fold cross-validation model proposed to detect CAT was evaluated using the following statistical classification measures: precision, sensitivity, specificity, positive predictive value (PPV), and negative predictive value. (NPV), F1 Receiver operating characteristic curve (AUC) score and area. AUC was calculated considering the corresponding probability of each predicted label. Precision, Sensitivity, Specificity, PPV, NPV, F1 Scores were calculated considering the most likely label predicted by the model. The manipulation threshold for considering labels as predictions of the model was set to 0.5.
ethical approval
This study was approved by the Human Research Ethics Committee of Maule Health Service and the Biomedical Research Ethics Committee of Talca Regional Hospital, Chile.
