Hybrid image enhancement with multi-stage DeepLabV3+ segmentation framework for scar detection in LGE-CMRI images

Teks telah dijajarkan (justified) dan dirapikan perenggannya mengikut standard format penulisan akademik untuk muat naik ke Institutional Repository (IR) UiTM, dengan mengekalkan 100% perkataan, huruf, tanda baca, dan simbol asal. Accurate myocardial scar segmentation from Late Gadolinium-Enhanced Cardiac Magnetic Resonance Imaging (LGE-CMRI) is essential for the assessment of myocardial infarction and treatment planning; however, it remains a challenging task due to low contrast between tissues, intensity inhomogeneity, blurry images, variation cardiac anatomy and size and complexity of scar regions. To address these limitations, this study proposed a comprehensive framework for automated myocardial structure and scar segmentation, beginning with the acquisition of LGE-CMRI datasets under approved ethical approval to ensure compliance with patient confidentiality and research governance requirements. Following data collection, image pre-processing was conducted, consisting of image augmentation and image labelling, where augmentation strategies were applied to increase dataset variability and improve model robustness, while precise manual labelling established reliable ground truth masks for the left ventricle, cavity, myocardium, and scar regions. After pre-processing, the images underwent image enhancement, incorporating both contrast improvement and image sharpening to increase visibility of subtle myocardial tissues and enhance structural boundaries without compromising anatomical integrity. Building upon this stage, a new Hybrid Contrast Enhancement–Image Sharpening (Hybrid CE-IS) approach was implemented to further refine intensity transitions and structural clarity, thereby generating more discriminative features for segmentation. The enhanced images were subsequently processed using a multi-stage DeepLabV3+ segmentation framework, where the first stage captured the anatomical structures of the left ventricle and cavity, the second stage refined myocardial tissue representation, and the third stage focused on precise delineation of small and irregular scar regions. To maximize segmentation effectiveness, a proposed Hybrid Image Enhancement–Image Segmentation (Hybrid IE-IS) strategy integrated the outputs of the Hybrid CE-IS with the multi-stage segmentation architecture, enabling improved feature extraction and spatial consistency across all myocardial structures. Performance evaluation demonstrated that the hybrid IE-IS framework achieved mean Dice Similarity Coefficient values of 98.22% for the left ventricle, 97.78% for the cavity, 85.15% for the myocardium, and 70.88% for myocardial scar, with corresponding Intersection over Union values of 91.67%, 93.75%, 75.68%, and 62.37%, and mean accuracy values of 96.57%, 95.85%, 83.54%, and 65.21%, respectively. In comparison, the multi-stage framework without hybrid enhancement achieved mean Dice values of 96.98%, 95.48%, 83.32%, and 68.49% for the respective regions, indicating that the hybrid enhancement contributed improvements of approximately 1–3% across all targets. These findings confirm that integrating structured data augmentation, hybrid image enhancement, and a progressive multi-stage segmentation strategy effectively improves feature representation and segmentation reliability, offering a robust and clinically applicable approach for automated myocardial scar analysis and decision support in cardiac assessment.