Expedited Colorectal Cancer Detection Through a Dexterous Hybrid CADx System With Enhanced Image Processing and Augmented Polyp Visualization

Abstract

The complexity and variability of medical imaging continue to impair the feasibility of early detection of colorectal cancer, despite its critical role in improving patient outcomes. This research presents a new multistage ensemble method that combines the strengths of three advanced deep learning models: vision transformers, RDV-22 (a mix of ResNet-50, DenseNet-201, and VGG-16), and Bidirectional Long Short-Term Memory (BiLSTM) networks. It improves the accuracy and robustness of Colorectal Cancer (CRC) detection. For suitable evaluation of our methodology, we adopted two benchmark datasets: CVC Clinic DB for binary classification and Kvasir dataset for multiclass classifications. To decrease the sizes of data significantly along with the generalizability of the model, we have used various data augmentation techniques, including magnification, inversion, and rotation. The best performing model was RDV-22+ BiLSTM + Vision Transformers, which recorded a 95.0% test accuracy and an area of 0.96 under the curve with the CVC Clinic DB dataset. The model showed a high performance in wide classes of samples such as esophagitis, polyps, and ulcerative colitis, with an accuracy rate of 92.5% and Area Under Curve (AUC) values 1.00 on the Kvasir dataset. We applied Local Interpretable Model-Agnostic Explanations (LIME) visualizations to both datasets, which proved to be an insightful identification of regions that the model relied on during the polyp detection process. Based on the LIME explanation, this analysis found that the visualizations generated by the Kvasir dataset were more informative and even clearer when complex cases, including ulcerative colitis and polyps are concerned. This indicates the model’s ability to handle challenging medical images effectively. Because of the complexity and diversity of the datasets, these results constitute a significant improvement on the current methodology. There is expectation that the proposed ensemble method could significantly improve the dependability and precision of diagnosing colorectal cancer. More accurate diagnoses and faster diagnoses in clinical practice might improve patient outcomes and reduce death rates if this is successful. While the ensemble model presents highly improved performance, this computationally and dataset-specific performance requirement suggests further optimization and validation for larger clinical applicability.

Keywords

• Enter colorectal cancer detection
• deep learning ensemble
• RDV-22
• BiLSTM
• vision transformers
• LIME visualization