Automatic recognition of left ventricular structures and cardiac cycles in 2D echocardiography images using deep reinforcement learning

Understanding the anatomical structures on the left side of the heart, particularly the Left Atrium (LA) and Left Ventricle (LV), including the Endocardium (LV Endo) and Epicardium (LV Epi), is essential for evaluating cardiac function. Manual segmentation of these structures in echocardiography remains the standard practice but is time- consuming, operator-dependent, and prone to variability. Developing automatic segmentation and identification methods for echocardiography images is crucial to enhance diagnostic accuracy, reduce subjectivity, and streamline clinical workflows. This thesis presents a novel Deep Reinforcement Learning (DRL) framework designed to address the challenges in segmenting cardiac structures in 2D echocardiography images. The proposed framework integrates a Convolutional Neural Network (CNN) for the identification of cardiac cycles (End-Diastolic (ED) and End-Systolic (ES) phases) with the DRL approach for precise segmentation of LV Endocardium, LV Epicardium, and LA. Unlike conventional Deep Learning (DL) techniques that rely on static, pre-labelled datasets, DRL dynamically interacts with the environment, enabling adaptive learning of complex spatial and temporal patterns inherent in Transthoracic Echocardiography (TTE). This adaptability makes DRL particularly effective in handling variations and complexities associated with TTE imaging, where external factors often degrade image quality. The research makes several key contributions to the field of echocardiography image analysis which are, A novel integration of a fully connected neural network as an actor in the Actor-Critic (A-C) framework to enhance the precision of capturing actions and improve segmentation accuracy. Development of a robust DRL algorithm trained iteratively on the CAMUS dataset, a publicly available dataset of 450 clinical cases, to achieve accurate segmentation without manual intervention. And A tailored CNN-based approach to identify ED and ES cardiac cycles, enabling phase-specific segmentation and overcoming ambiguities caused by the similarity of ED and ES images.The DRL framework was rigorously evaluated using metrics such as Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), Jaccard Index, and Mean Absolute Error (MAE), demonstrating superior performance compared to traditional DL methods. Results highlight the model’s capability to adapt to varying imaging conditions and its potential to transform clinical cardiology by providing a scalable, accurate, and efficient solution for cardiac structure recognition. This thesis addresses a critical gap in the application of DRL to medical imaging, setting a foundation for further exploration and development in this domain.

Text FSKTM 2024 22 - Full Text.pdf Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (7MB) Text FSKTM 2024 22.pdf Restricted to Repository staff only Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (8MB)

Actions (login required)

View Item