Deep learning-based calcium scoring of the aortic valve using 3D TEE: Preliminary study

• Main Author: Bairros, Rita Seixas
• Publication Date: 2024
• Format: Master thesis
• Language: eng
• Source: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Download full: http://hdl.handle.net/10071/32919
• topic: Echocardiography; Aortic stenosis; Calcium scoring; Deep learning; Processamento de imagens -- Image processing; Ecocardiografia; Estenose aórtica; Quantificação de cálcio

Aortic stenosis is a critical cardiovascular condition that can be assessed through echocardiography, with calcium deposits on the aortic valve playing a key role in diagnosis. This dissertation presents a hybrid approach combining deep learning and image processing methods to improve the detection and quantification of aortic valve calcifications. Two main objectives were addressed: (1) detecting and extracting the image region corresponding to the aortic valve, and (2) quantifying calcium deposits within the segmented valve, correlating these results with Agatston scores derived from CT scans. An adapted YOLOv8n model was employed for valve detection, achieving 99.94% precision, 81.82% recall, and a mean Average Precision (mAP) of 92.88%. The region of interest was successfully extracted in all cases using a combination of manual annotations and automated segmentation techniques. For calcium scoring, two approaches were explored: a heuristic method and convolutional neural network (CNN) models. The CNN models captured complex patterns in the echocardiographic images, with the fine-tuned ResNet50 model demonstrating superior performance, achieving a mean absolute error of 1356.56. The heuristic method showed a Pearson correlation of 0.75 with the CT-derived Agatston score, validating its accuracy, especially in patients with higher calcium scores. Additionally, a gender-based analysis revealed that male patients exhibited higher calcium deposits, consistent with existing cardiovascular research. This work shows that combining deep learning with traditional methods can improve the diagnostic process for aortic stenosis, offering potential for timely, precise diagnoses and advancing healthcare system efficiency.