A fully automated myocardial infarction segmentation and detection in late gadolinium enhancement cardiac magnetic resonance imaging using deep learning

Late gadolinium enhancement cardiac magnetic resonance imaging (LGE-CMR) has emerged as the gold standard for non-invasive myocardial tissue characterization, particularly for identifying scarred or infarcted myocardial regions. The current diagnostic workflow remains reliant on manual segmentation and expert interpretation, which are time-consuming and subject to inter-observer variability. Although deep convolutional neural networks (DCNNs) have demonstrated strong potential in automated medical image segmentation, existing models suffer from low segmentation accuracy due to the intrinsic complexity of the modality. This study proposes a fully automated, end-to-end framework for MI detection, comprising two main components: a segmentation of myocardial scars and a subsequent detection module that employs sequential analysis to identify MI cases based on the segmented regions. As a foundational step, this study proposes a new automated classification method to address a gap in previous studies that relied on manual slice preselection. A shallow deep convolutional neural network (S-DCNN) model is introduced to automatically classify LGE-CMR slices into those containing the left ventricle (LV) and non-LV categories. This classification step significantly streamlines the automated segmentation pipeline by ensuring that only relevant slices are processed further. Central to the proposed approach is a novel dual-stage segmentation model designed for fully automated myocardial infarction detection from LGE-CMR images. This model comprises two task-specific stages, each built upon a customized DeepLabV3+ segmentation network. The first stage, referred to as DeepLabV3+ Tailor LV (DLT-LV), is tailored for accurate segmentation of the LV region. The second stage, DeepLabV3+ Tailor Scar (DLT-Scar), focuses on segmenting myocardial scar tissue within the segmented LV area. To improve the reliability of this pipeline, a morphological post-processing procedure is introduced between the two stages to refine the LV segmentation output and address common issues such as small gaps and isolated mis-segmentations. To further enhance the performance of the second stage, a hybrid optimization-segmentation approach is proposed. This method integrates Particle Swarm Optimization (PSO) into the DLT-Scar model by automatically tuning the hyperparameters of the Tversky loss function used in its classification layer, specifically addressing the class imbalance inherent in scar segmentation. Finally, in the MI detection phase, a Scar Sequential Slice Reconstruction (3SR) module is integrated with a S-DCNN classification model to exploit the spatial continuity of myocardial scars across consecutive slices. This module constructs a localized sequence of consecutive image slices with the largest segmented scar identified during the segmentation stage. During the performance evaluation stage, the results were compared with patient radiology reports to assess their clinical relevance. Experimental results demonstrate that the proposed method achieves high segmentation accuracy and robust MI detection, with a mean Dice Score of 74.31% for scar segmentation and 96.7% accuracy, sensitivity of 100% and specificity of 95% for MI classification. These findings offer a scalable and fully automated solution for improving diagnostic consistency and efficiency in clinical cardiac imaging workflows.