OVT-Net: Semantic segmentation of gastrointestinal cancer using an optimized vision transformer model with explainable AI
Year: 2025
Type: article
Abstract: Gastric and intestinal cancers are among the deadliest gastrointestinal diseases, necessitating precise organ segmentation for effective detection and treatment planning. Conventional deep-learning models, such as CNN-based U-Net architectures, struggle with long-range dependencies and complex anatomical variations. This study introduces OVT-Net (Optimized Vision Transformer Network), an innovative deep-learning framework integrating Swin Transformer blocks, EfficientNetB7, Adaptive Contextual Attention (ACA) module, Atrous Spatial Pyramid Fusion (ASPF), and Squeeze-and-Excite (SE) Networks. Unlike traditional architectures, OVT-Net employs a hybrid dual-encoder structure, combining EfficientNetB7 for low-level feature extraction and Swin Transformers for global context modeling, addressing intricate anatomical complexities and imaging variabilities of the gastrointestinal tract. The model is trained on 38,496 MRI/CT scans paired with RLE-encoded masks that contain structural and labeling inconsistencies. These inconsistencies are resolved through a comprehensive preprocessing pipeline incorporating path generation, label restructuring, and augmentation to improve generalizability. Experimental results demonstrate superior performance, with a Dice score of 0.9350, an IoU score of 0.9218, a BCE loss of 0.0716, and robust surface distance metrics (HD95 and ASSD), outperforming conventional segmentation methods. To enhance clinical applicability, Explainable AI (XAI) techniques, including Grad-CAM and Grad-CAM++, provide interpretability by highlighting critical regions, improving model transparency in decision-making. Furthermore, OVT-Net is deployed in a Django-based web application, facilitating real-time segmentation and classification with an average accuracy of 97.5 %. This research presents OVT-Net as a transformative AI-driven segmentation model, bridging advanced vision transformers with XAI for enhanced medical diagnostics. Its integration into real-world clinical settings paves the way for improved cancer detection and early intervention.
Source: Clinical eHealth
Authors: Anika Tahsin Meem, Sifat Momen
Institution: North South University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $995
OpenAlex ID: W4416684116