AI radiomics pipeline achieves 0.80 AUC for endometriosis MRI subtyping

A team of researchers from Nantes University (LTSI, EPUN) has developed a radiomics-based pipeline to subtype endometriosis using pelvic MRI scans from the publicly available UT-EndoMRI dataset. The approach extracts radiomics features from manually segmented uterine and ovarian regions, comparing multi-scale feature representations and feature-selection strategies.

Using raw Wavelet-derived features with a Gradient Boosting classifier, the team achieved a peak AUC of 0.80 in distinguishing patients with endometriomas from those without. However, the model suffered from high false positives, resulting in low specificity. The study also found that ComBat harmonization—a technique to reduce batch effects—did not consistently improve performance, particularly in small, multi-site datasets where acquisition groups contained few patients. Unsupervised clustering identified reproducible but poorly separated partitions linked to acquisition variables, underscoring the fragility of radiomics-based subtyping in heterogeneous datasets.