Document Type
Abstract
Publication Date
2-11-2026
Abstract
Introduction— Current noninvasive assessment of nonalcoholic fatty liver disease (NAFLD) relies primarily on MRI-based fat quantification, such as proton density fat fraction (PDFF). This study explored the use of an MRI radiomics–based machine learning approach for accurate patient-level classification of NAFLD.
Methods— Liver MRI images from patients with suspected NAFLfD were retrospectively analyzed. Radiomics features were extracted using PyRadiomics from full-liver masks across 4–6 axial slices per patient. Slice-level features were aggregated to the patient level using summary statistics (mean, median, and standard deviation). To reduce redundancy and multicollinearity, features were pruned using Pearson correlation analysis with a threshold of |r| > 0.95. A two-stage feature selection strategy was employed to meet model constraints: correlation-based pruning followed by SelectKBest using the ANOVA Fstatistic. The final feature set consisted of 96 radiomics features and 4 clinical metadata variables (age, BMI, sex, and MRI PDFF fat percentage). A TabPFN classifier was trained and evaluated using a patient-level train/test split.
Results— Feature pruning reduced the radiomics feature space from 306 to 133 features (56% reduction). The final TabPFN model achieved an accuracy of 95.65% and an area under the receiver operating characteristic curve (AUC) of 0.992 on an independent test set of 23 patients (11 non-fatty, 12 fatty).
Conclusions— An MRI radiomics–based machine learning pipeline using TabPFN demonstrated excellent performance for patient-level classification of NAFLD. This approach highlights the potential of radiomics combined with modern probabilistic foundation models to enhance noninvasive liver disease assessment and may complement conventional clinical workflows.
Recommended Citation
Kirichenko, Vladimir and Vadigepalli, Rajanikanth, "Noninvasive Patient-Level Detection of NAFLD Using MRI Radiomics and Machine Learning" (2026). Phase 1. Paper 4.
https://jdc.jefferson.edu/si_dh_2028_phase1/4
Language
English
Comments
Presented at the 2026 Scholarly Inquiry (SI) Research Project Symposium.