A Residual Learning Approach Towards the Diagnosis of Colorectal Disease Effectively

Colorectal cancer (CRC) is a significant global health concern with substantial morbidity and mortality rates. Early identification and proper diagnosis significantly impact the prognosis of patients with CRC. This study presents an efficient residual learning model using ResNet50 as the underlying architecture with a comprehensive analysis focused on developing a robust computer-aided diagnostic (CAD) system for CRC detection and classification using colonoscopy images. This approach involves extensive preprocessing, data augmentation and hyperparameters tuning techniques to enhance the model’s ability to detect subtle abnormalities and overcome potential data deficiency. Furthermore, dropout layers and regularization techniques are introduced to address overfitting issues, improving the model’s generalization performance. The study uses four different benchmark datasets: PolypsSet, CP-CHILD-A, CP-CHILD-B, and Kvasir v2. These datasets contain diverse colonoscopy images, including polyps, non-polyps, adenomas, and hyperplastic lesions. The model exhibits remarkable Accuracy, Precision, Recall, and F1-score, effectively on all four datasets. The Recall of the proposed model across all datasets surpassed 87%, which outperformed three benchmark results out of four, validating its ability to effectively classify colorectal polyps. Consequently, this study advances the field of medical image analysis and opens the door to early detection and individualised treatment plans for patients with CRC.