Comparative Analysis of Optimized Logistic Regression with State-of-the-Art Models for Complex Gastroenterological Image Analysis

Daniela-Maria Cristea¹Ioan Sima²Laszlo Barna Iantovics^1*

• ¹George Emil Palade University of Medicine, Pharmacy, Sciences and Technology of Târgu Mureş, Târgu Mures, Romania
• ²Universitatea Babes-Bolyai, Cluj-Napoca, Romania

To classify gastrointestinal (GI) polyps detected in colonoscopy images, an essential task in colorectal cancer prevention, it was optimised Multiclass Logistic Regression (LR), a preferred machine learning (ML) algorithm among physicians due to its interpretability. Given the diagnostic ambiguity of serrated polyps, which share features with both hyperplastic and adenomatous lesions, we focused on multiclass classification using a structured dataset of 152 instances and 698 extracted features. A key objective was to explore hyperparameter tuning approaches for ML models. We proposed a decision rule to guide model selection and conducted a statistical analysis of 88 distinct LR configurations, varying solvers, penalties, and regularisation strengths. The best-performing LR model (liblinear solver, L1 penalty, C = 0.01) achieved 70.39% accuracy, exceeding the average performance of physician evaluators (experts: 65.00%, beginners: 58.42%). To enhance classification performance, four additional ML algorithms were implemented and benchmarked: k-Nearest Neighbours (kNN), Support Vector Machine (SVM), Random Forest (RF), and XGBoost. For each classifier, parameter tuning was applied using grid search and performed stratified cross-validation. In the multiclass setting, XGBoost achieved the highest macro-average F1-score (0.88) and overall accuracy (89%), followed by Random Forest (F1 = 0.85, accuracy = 86%), SVM (F1 = 0.83, accuracy = 84%), and kNN (F1 = 0.56, accuracy = 66%).