COLON POLYP IDENTIFICATION AND CLASSIFICATION — EXPLORATORY STUDY OF MULTIVARIATE FEATURE DRIVEN OPTIMIZED HETEROGENEOUS CLASSIFIER

Colon cancer, also known as colorectal cancer (CRC), initiates with the formation of benign growths termed polyps, marking a crucial stage in its progression. While colonoscopy remains the gold standard for CRC screening, manual inspection presents inherent limitations, including the risk of overlooking polyps due to their resemblance to surrounding mucosa. Challenges also arise from variations in endoscopist expertise and focus. In this study, we propose an intelligent approach for automated detection and classification of polyp image frames. Our method harnesses the advanced multivariate feature extraction techniques to effectively address these complexities. In this study, we adopt a filter-based approach for feature selection, utilizing Mutual Information (MI) to discern and prioritize features based on their information-theoretic relevance to the target variable. Principal Component Analysis (PCA) is applied to the selected features to reduce the feature space, effectively handle multicollinearity, and mitigate the curse of dimensionality inherent in medical data, ultimately leading to better performance of machine learning models and potentially preventing overfitting. We explore traditional machine learning approaches to distinguish polyp image frames, utilizing heterogeneous ensemble classification models trained with hyperparameter optimization through the Treestructured Parzen Estimator (TPE). Our study includes representative images of both polyp and non-polyp cases, covering healthy tissues and other pathologies. The model achieved impressive performance in polyp frame identification on same dataset, with a precision of 93.89%, recall of 93.79%, and F1-score of 93.99%. Additionally, it attained an accuracy of 93.89% and a high ROC AUC of 0.9873, demonstrating its effectiveness in accurately identifying polyp images. To assess generalizability, the model was evaluated on the Kvasir-v2 dataset, which was not used during training. It maintained strong performance, achieving a mean precision of 85.75%, mean recall of 86.51%, and mean F1-score of 86.11%, indicating good robustness across datasets.