چكيده به لاتين
According to the World Health Organization, cardiovascular diseases are the leading causes of death in the world. Cardiac ultrasound, also known as echocardiography (or echo), is the standard method for screening, diagnosing, and monitoring cardiovascular diseases. This non-invasive imaging modality has gained a special place in diagnosis and treatment as the most economical and noninvasive method. Echo-standard examinations include evaluation of the size and function of the cavities, as well as stenosis and other physical, anatomical, and movement features of the valves, walls, and cavities of the heart. Interpretation of echocardiographic images is difficult even for the best specialists due to their noisy nature. Also, many measurements and analyzes to diagnose heart function are time consuming and depends on the experience of the echocardiographer. Therefore, there is a great need for an automated system to improve image quality and extract the main parameters of heart function to help with diagnosis.
Deep learning methods such as convolutional neural networks (CNNs) and transformers have yielded good results in many AI problems. With the help of deep learning, part of the responsibility of extracting features has been transferred from human to computer. The unprecedented success of deep learning is largely due to the following factors: (1) progress in GPUs; and (2) availability of large amounts of data (such as big data).
Deep learning has played a major role in the analysis of medical images in modalities such as MRI, CT, and X-ray. The use of this type of learning is in one of the applications of medical image processing, classification and labeling of objects and images with the approach of computer prediction or diagnosis (CADx). In this dissertation we seek to apply the idea of deep learning in feature extraction from echocardiographic images and classification of images based on the type of disease (CAD systems), extraction of information from echocardiographic images during the cardiac cycle and the use of memory networks.
According to the literature review, not much research has been done on the application of the idea of deep learning in the analysis of echocardiographic images. Therefore, the aim of this dissertation is to apply the idea of deep learning in order to classify the morphology of the mitral valve. To achieve this goal, standard data (transthoracic echocardiogram (TTE) images) with appropriate diversity during two years were collected.
These data consist of two parts: prosthetic mitral valve dataset (447 Prostatic and 1597 natural) and Carpentier mitral valve dataset (424 NL, 155 MVP, 392 IIIa and 802 IIIb). In the first dataset, using 13 common CNN-based networks to classify mitral valve echocardiographic images in both natural and prosthetic classes, the EfficientNetB3 network in A4C view and the EfficientNetB4 network in PLA view had the best performance with AUC = 0.99. For the second dataset, the proposed algorithm framework has three parts of automatic analysis: view classification with EfficientNetB0 network, left ventricular location detection with RetinaNet network to determine systole-diastole phases and classification of 4th and 3rd classes. After analyzing the first two sections, the PLA view of each sample is first assigned by the 4-class Inception-ResNet-v2 network to classify MVP disease. If the label is non-MVP, the A4C view is checked by the 3-class ResNeXt50 network to determine one of the three classes IIIa, IIIb and NL. With such a rule-based framework, we achieved 80% accuracy in the test data. As an additional contribution, to provide explainability, we applied the Gradient-weighted Class Activation Mapping (Grad-CAM) algorithm to visualize areas of echocardiographic images where the different models generated a prediction.
By using new idea of transformers and the mechanism of attention, we evaluated the use of interframe information in the aforementioned four classes. Using transformers to have equal conditions in a cardiac cycle, we first upsampled the PLA view frames of all samples by SPARTEMP method; 16 samples are then downsampled from the frames of a cycle and then the feature vector is extracted from each frame by the Inception-ResNet-v2 network. These vectors are pre-processed into self-attention blocks. In these blocks, classification was performed for each of the 4 classes by generating key vectors, value and search and applying nonlinear operator, convolution block and mean integration. With this method, we achieved a 2% increase in accuracy compared to the previous method. The results of this study can be used as a computer-aided diagnosis system (CADx) in the software of echocardiography devices.
