Comparative analysis of optimized logistic regression with state-of-the-art models for complex gastroenterological...

Area of Science:

• Gastroenterology
• Medical Imaging
• Machine Learning

Background:

• Accurate classification of gastrointestinal (GI) polyps is crucial for colorectal cancer prevention.
• Serrated polyps present diagnostic challenges due to morphological similarities with hyperplastic and adenomatous lesions.

Purpose of the Study:

• To evaluate machine learning (ML) techniques for multiclass classification of GI polyps from colonoscopy images.
• To compare the performance of Logistic Regression (LR) against other ML algorithms.

Main Methods:

• Optimized Logistic Regression (LR) by analyzing 88 configurations.
• Implemented and tuned k-Nearest Neighbors (kNN), Support Vector Machine (SVM), Random Forest (RF), and XGBoost using grid search and cross-validation.
• Utilized a dataset of 152 instances with 698 features for training and evaluation.

Main Results:

• The best LR model achieved 70.39% accuracy, surpassing physician benchmarks.
• XGBoost demonstrated the highest performance with a macro-average F1-score of 0.88 and 90% accuracy.
• Random Forest and SVM also showed strong results, outperforming kNN.

Conclusions:

• While Logistic Regression (LR) offers interpretability, ensemble methods like XGBoost and Random Forest provide superior accuracy and robustness for GI polyp classification.
• These advanced ML models are suitable for clinical decision support, especially in data-limited situations.