A Hybrid Framework for Enhanced Segmentation and Classification of Colorectal Cancer Histopathology

Aaseegha M DVENKATARAMANA B^*

• Vellore Institute of Technology, Vellore, India

Colorectal cancer (CRC) is one of the leading causes of cancer-related deaths worldwide. The results for patients are improved by early detection and precise diagnosis. Manual inspection of stained tissue slides is essential for conventional histological evaluation, but it is time-consuming, prone to observer variability, and susceptible to inconsistent diagnoses. Although there is potential in deep learning architectures like U-Net, ResNet, and traditional CNN-based models, these models often face challenges such as poor feature extraction, inadequate global context modelling, and incorrect classification of tissues with similar morphologies. To address these problems, we propose a hybrid deep learning system that combines Swin Transformer, EfficientNet, and ResUNet-A. This model uses self-attention, compound scaling, and residual learning to improve categorization, global contextual modelling, and spatial feature extraction. For training and evaluation, a histopathological dataset that included serrated adenoma, polyps, adenocarcinoma, high-grade and low-grade intraepithelial neoplasia, and normal tissues was utilized. With 93% accuracy, 92% precision, 93% recall, and 93% F1-score, the hybrid model outperformed individual architectures in segmentation and classification. Expert annotations and segmentation masks closely matched, demonstrating the method's dependability. This hybrid design offers a reliable instrument for the automated analysis of histopathological features in colorectal cancer.