Enhanced colorectal gland segmentation through multi-scale attention and contextual feature fusion.

Histological image segmentation of colorectal cancer plays a crucial role in reliable cancer grading, prognosis evaluation, and treatment planning. Nevertheless, existing automated approaches often struggle due to fine-grained glandular structures, staining variations, acquisition differences across scanners and highly heterogeneous tissue morphology. To address these challenges, we propose MAC-Net, an enhanced deep learning model that integrates multi-scale feature fusion with attention-guided contextual decoding to achieve precise gland segmentation across diverse tumor differentiation stages. MAC-Net retains fine structural information by channel-wise attention, enhances learning of discriminative features by the increased number of enriched encoder-decoder lateral connections, and the global contextual information, by means of multi-scale spatial pooling at the bottleneck. To determine the generalizability of the model, we trained the model on the EBHI-Seg data (2228 images) and cross-validated it on the GIaS data (165 images). MAC-Net achieved 95.08% Dice, 90.92% IoU, 95.83% Precision, and 95.65% Recall, outperforming existing architectures. Furthermore, Gradient-weighted Class Activation Mapping visualization and uncertainty analysis enhance interpretability, provides deeper insight on model's transparent behavior and supports reliable clinical decision-making in the field of digital pathology.