Advanced colon cancer detection: Integrating context-aware multi-image fusion (Camif) in a multi-stage framework

Colon cancer begins in the large intestine, often evolving from benign polyps into malignant cancer. Early detection through screening is vital for effective treatment and better survival rates. Risk factors include age, genetics, diet, and lifestyle, with symptoms like changes in bowel habits and blood in the stool, though early stages may be asymptomatic. This work proposed a comprehensive multi classes detection and classification of colon cancer. In this work we used publicly available Curated Colon Dataset to diagnose conditions such as esophagitis, ulcerative colitis, polyps, and normal cases. The proposed approach uses advanced deep learning models to integrate pre-processing, segmentation, and classification. The process begins with pre-processing, which involves resizing, contrast enhancement, noise reduction, and normalization of pixel values. This work proposes a Context-Aware Multi-Image Fusion (CA-MIF) technique in the preprocessing phase to improve the visibility of blood vessels and tissue texture, enhancing diagnostic accuracy. The processed images are then input to a U-Net++ model for segmentation, generating masks highlighting regions of interest, including the colon and affected areas. Post-segmentation, image enhancement techniques further refine the quality and clarity of the images. Enhanced images are then classified using the ResNet-50 model, trained to categorize images into four distinct classes: esophagitis, ulcerative colitis, polyps, and normal. In the classification phase, cancerous classes (ulcerative colitis and polyps) undergo additional segmentation using DeepLabv3+. Model 1 (DeepLabv3+) is applied to ulcerative colitis, generating detailed masks to analyze affected regions, while Model 2 (DeepLabv3+) is used for polyps. For the U-Net++ and DeepLabv3+ models, evaluation measures are segmentation accuracy, precision, recall, and F1 score; for the ResNet-50 model, these metrics are classification accuracy, precision, recall, and F1 score. When it comes to detecting and differentiating between malignant and non-cancerous illnesses, the framework achieves great accuracy., demonstrating its effectiveness and potential for clinical applications in medical image analysis. The results indicate the proposed method’s high efficacy, achieving an F1 score of 99.31. It also demonstrated strong performance metrics with a specificity of 99.91, sensitivity of 99.10, accuracy of 98.18, and a Dice coefficient of 99.82, highlighting its robust capability in accurately detecting colon cancer.