Abstract
Clinical applications of machine learning in prostate MRI are often plagued by domain shifts caused by acquisition parameters, manufacturer-specific processing, and magnetic field strength (MFS). The present study evaluates the efficacy of two bias mitigation techniques: (i) a customized bias mitigation strategy that combines synthetic data generated via the Bayesian Gaussian Mixture Model with optimal component estimation (BGMMOCE) along with Gower distance-based matching and graph-driven community selection to balance underrepresented groups, and (ii) fair reweighing using the IBM AIF360 frame-work. Both approaches were evaluated on prostate lesion detection (UC1) and cancer aggressiveness (UC2) prediction by utilizing T2-weighted radiomics derived from both prostate and lesion volumes. Bias mitigation was assessed against the baseline raw radiomics using a rigorous experimental protocol with training sets adjusted via fair reweighing or generative fairness, while testing sets were stratified for both clinical endpoints and biases (manufacturer and MFS) to ensure all classes are represented for model evaluation. The performance gaps between privileged and under-privileged groups were quantified. Mitigation strategies yielded significant improvements in lesion-level tasks (up to +3.7% AUC), demonstrating that fairness interventions can simultaneously enhance downstream performance in signal-rich settings. Conversely, in tasks with low intrinsic predictive signals (gland-level classification), fairness adjustments led to minor performance degradation. Active bias mitigation shows strong potential for improving subgroup equity, though its impact remains highly dependent on the baseline target signal and anatomical feature granularity. In conclusion, active bias mitigation in radiomics has the potential of improving model generalizability. In this study it is demonstrated that for clinical applications, bias mitigation strategies not only reduce distribution disparities regarding image acquisition bias but could also enhance the overall predictive power of a model.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/CaPTion_011.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/profile?id=~Eleftherios_Trivizakis1
BibTex
@InProceedings{TriEle_GraphRefined_MICCAISAT2026,
author = { Trivizakis, Eleftherios AND Pezoulas, Vasileios C. AND Tachos, Nikolaos AND Tsiknakis, Manolis AND Fotiadis, Dimitrios I. AND Regge, Daniele AND Papanikolaou, Nikolaos AND Procancer-I Consortium AND Marias, Kostas},
title = { { Graph-Refined Probabilistic Mitigation and Fair Reweighing for Enhanced Equity and Generalizability in Prostate MRI Radiomics } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17278},
month = {pending},
page = {pending}
}