Deep learning-based aortic valve region of interest and calcification detection in echo images

The automatic detection of heart diseases is currently of widespread interest as a tool to enhance the accuracy of cardiologists' diagnoses and reduce diagnosis time using echocardiography images. Aortic stenosis is a fundamental condition in heart disease and can lead to various other health issues if it progresses. The formation of calcified areas in the aortic region of the heart is a precursor to aortic stenosis. Therefore, detecting abnormalities in the aortic region of the heart is crucial for diagnosing and treating the disease. Although high-resolution computed tomography (CT) imaging is generally preferred for detecting heart diseases, disadvantages such as cost and patient radiation exposure necessitate alternative approaches. In this context, utilizing ultrasound imaging to reduce radiation risks and maintain cost-effectiveness becomes crucial. In the scope of this thesis, fully automatic detection of the aortic region of interest and the calcified areas within that region on echocardiography images obtained from heart ultrasound was performed without the need for CT imaging. A new lightweight model designed for automatic aortic valve detection, AVD-YOLOv5, is proposed. This model includes various enhancements to the YOLOv5 architecture. Notably, the use of depthwise separable convolution significantly contributes to the lightweight design of the model by reducing the number of parameters while maintaining precision. Additionally, a new and large dataset consisting of 320 echocardiography images specifically for aortic valve detection was created. Experimental results show that the modified ADV-YOLOv5 model achieves an accuracy of 94.3% and a precision of 86.8%. The model also demonstrates a significant reduction of 67% in inference time compared to the original YOLOv5 model. Despite a marginal decrease of 0.94% in precision, the model's efficiency has significantly improved. As the next step for predicting the aortic valve calcium score, a fully convolutional neural network approach is proposed for segmenting the calcified areas in the aortic valve from echocardiography images. This aims to automatically calculate the aortic valve calcium score more objectively, quickly, easily, cost-effectively, and without radiation. The proposed method achieved 86.3% accuracy, 80.27% precision, 76.27% Dice coefficient, and 61.9% Jaccard index on a new dataset consisting of 118 parasternal short-axis echocardiography images from 82 different patients. These results demonstrate the feasibility and potential effectiveness of fully automating the aortic valve calcium scoring from echocardiography images. The proposed system can be used by cardiologists for more efficient and reliable diagnosis.