A Disentangled Representation-Based Multimodal Fusion Framework to Detect Colorectal Cancer

Background: Computer-aided segmentation is an important activity that supports gastroenterologists in assessing and eradicating abnormal tissue from the gastrointestinal system. Diseased polyps mostly arise in the colorectal region of the gastrointestinal system and in the mucous membrane. They include micro-abnormal tissue protrusions that increase the risk of incurable illnesses such as cancer. Hence, early evaluation of polyps can reduce the likelihood of their progression into cancer, such as adenomas, that can develop into colorectal cancer. Methods: Deep learning-based diagnostic tools play a critical role in the early detection of illnesses. To segregate polyps using colonoscopy frames, a deep learning approach known as Mask R-CNN is presented. Mask Region based Convolution Neural Network with an Region Proposal Network backbone is a modified version of the standard Faster-Convolution Neural Network, comprising three stages: feature extraction, region of interest (ROI) pooling, and segmentation for severity analysis. The mask R-CNN technique outperforms previous deep learning techniques in segmentation trials using a single dataset obtained from the large intestine of the gastrointestinal system through colonoscopy. Results: The predicted technique exceeds having a mean Dice score of 98%, an accuracy of 96.6%, a precision of 97%, a recall of 96.6%, and an F1-score of 92%. Conclusion: The results highlight the potential of disentangled representation learning in addressing the heterogeneity of biomedical data and improving diagnostic performance. Moreover, the interpretability of the learned representations offers valuable clinical insights, fostering trust in artificial intelligence–assisted diagnostic systems.