Comparative evaluation of conventional Machine learning and automated pipeline optimization for MRI radiomics in Hepatocellular carcinoma

Background Hepatocellular carcinoma (HCC) is a prevalent cancer in Malaysia and often detected at advanced stage. Utilisation of automated machine learning (AutoML) with radiomics data reduce reliance on visual interpretation in clinical practice. Aim This study evaluated the performance of conventional Support Vector Machine (SVM) incorporated with Analysis of Variance (ANOVA) F-test and Binary Particle Swarm Optimization (BPSO) and investigated the effect of selected features on the AutoML for differentiating stages of HCC on magnetic resonance imaging (MRI). Methods A total of 662 radiomic features were extracted from the segmented HCC MR images. The proposed conventional model utilised selected features with an optimized SVM classifier. Four AutoML configurations, TPOT-Default, TPOT-LinearSVC, H2O, and AutoGluon were evaluated in two phases: Phase 1 employed the original feature set whereas Phase 2 used the selected subset. Performance was assessed for all nine models using accuracy, precision, recall, specificity, F1-score, area under curve (AUC) and precision-recall curve (PRC). DeLong and McNemar’s tests were employed for pairwise statistical comparisons. Results The conventional model achieved an accuracy of 0.650 and the highest numerical AUC (0.745) and PRC (0.836) among all evaluated models·H2O achieved the highest accuracy (0.700) in both phases. TPOT-LinearSVC showed the greatest improvement after implementation of feature selection, while AutoGluon improved in several threshold-dependent metrics. No pairwise differences were statistically significant (all p > 0.05). Conclusion The proposed feature selection strategy demonstrated a model-dependent effect on the classification performance. Although the conventional model achieved a high AUC and PRC, its performance was not statistically superior to the AutoML models.