Automated Cardiovascular Lesion Segmentation in Coronary CT Angiography Using Trans U Net: A Transformer-Based Deep Learning Approach

Abstract

Accurate and automated segmentation of cardiovascular lesions in Coronary CT Angiography (CCTA) is critical for early diagnosis and treatment planning of coronary artery disease (CAD). This study proposes a two-stage hybrid framework that integrates feature-based classification with deep learning-based segmentation for improved medical image interpretation. Initially, images undergo preprocessing, including resizing, contrast enhancement, and normalization, to enhance visual quality. Feature extraction is performed using Local Binary Patterns (LBP) and Adaptive Weighted Multi-Resolution Gray-Level Co-Occurrence Matrix (AWMR-GLCM), capturing texture and spatial characteristics. The extracted features are cascaded and classified using a Long Short-Term Memory (LSTM) network to learn temporal dependencies. Simultaneously, a deep learning-based segmentation model, Trans U Net, is trained to precisely delineate affected regions. The proposed method achieved superior performance, attaining 98.54% accuracy, 99.64% F1-score, and 99.04% precision for classification, while the segmentation stage obtained 98.25% DSC, 98.31% IoU, and 99.47% HD, demonstrating the robustness and reliability of the approach. Experimental results demonstrate the effectiveness of combining feature-based classification with deep learning-based segmentation, enhancing robustness and reliability in medical image analysis.