Colorectal cancer detection with enhanced precision using a hybrid supervised and unsupervised learning approach

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Colorectal cancer (CRC) detection and segmentation remain critical challenges in medical diagnostics.
• Existing methods often struggle with accuracy, interpretability, and precise localization of malignant regions.

Purpose of the Study:

• To introduce a novel hybrid ensemble framework for enhanced detection and segmentation of colorectal cancer.
• To improve the accuracy and understandability of diagnostic results in medical imaging.

Main Methods:

• Developed a hybrid ensemble framework integrating Convolutional Neural Network (CNN) models (ADa-22, AD-22), transformer networks, and Support Vector Machine (SVM) classifier.
• Utilized the CVC ClinicDB dataset comprising 1650 colonoscopy images.
• Incorporated K-means clustering with bounding box visualization for segmentation refinement.
• Employed hyperparameter optimization to balance performance and generalization, suppressing overfitting.

Main Results:

• The AD-22 + Transformer + SVM model achieved an AUC of 0.99, with 99.50% training and 99.00% testing accuracy.
• High accuracy for polyp (97.50%) and non-polyp (99.30%) detection, with excellent recall rates.
• K-means clustering yielded a silhouette score of 0.73, enhancing segmentation visualization.

Conclusions:

• The proposed hybrid framework effectively addresses limitations of previous approaches by combining CNN feature extraction, Transformer attention mechanisms, and SVM decision boundaries.
• The integration of unsupervised clustering improves segmentation visualization and overall diagnostic understandability.
• The framework demonstrates high performance and potential for clinical feasibility in colorectal cancer detection and segmentation.