Kardiyak mri ile koroner arterhastaliği siniflandirmasinin kendikendine gözetimli öğrenmeyoluyla geliştirilmesi

The study explores the implementation of Self-Supervised learning (SSL) in classifying Coronary Artery Disease (CAD) using Cardiac Magnetic Resonance Imaging (MRI). This study aims to improve diagnostic accuracy by harnessing SSL's potential to extract significant features from unlabeled data, which is crucial in medical imaging due to the difficulties associated with extensive manual labelling of data. The research encompasses the training of three models: two self-supervised models that implement transfer-learning, and a third model that employs a fully supervised approach, functioning as a reference point for evaluation. The pretext model undergoes training on unlabeled MRI data to develop strong representations, which are utilized in downstream tasks with varying levels of model modification. In the first model, every part of the pretext model remains fixed, retaining the learned features, while in the second model, alternating layers are allowed to train to improve task-specific characteristics. The third model depends entirely on labelled data for its training process. The results indicate that SSL methods enhance classification performance, particularly when contrasted with conventional supervised learning techniques. This study underscores SSL's capability to derive high-quality features from a limited number of labelled samples, thereby fortifying CAD detection systems against adversarial threats and out-of-distribution data. The comparative assessment of the models reveals SSL's potential to decrease dependence on labelled datasets while either maintaining or exceeding the accuracy levels of fully supervised models. Additionally, the research addresses the challenges and limitations associated with SSL, particularly model reliability, accuracy, and the necessity for a well structured pretext task to promote effective feature extraction. The results imply that SSL represents a promising strategy for improving CAD detection and paving the foundation for improved diagnostic procedures