Characterization and classification of deep endometriosis lesions in magnetic resonance imaging

Machine learning is being explored to improve the diagnosis of deep endometriosis. This study evaluates its potential to distinguish active from fibrotic lesions using radiomic and clinical data from 3D MRI scans. The dataset includes 3D MRI scans from 61 patients, where lesions were segmented, and radiomic features extracted from both 3D volumes and corresponding 2D slices. Clinical data were linked to each lesion based on the patient. Six datasets were constructed: radiomic-only (3D/2D), clinical-only (3D/2D), and combined features (3D/2D). Dimensionality reduction techniques (LDA, PCA, UMAP, PaCMAP) did not reveal a clear separation between lesion types, suggesting that the selected features lacked distinctive information. Classification was performed using the Tree-based Pipeline Optimization Tool (TPOT), optimizing pipelines for each dataset. The best performance was achieved with the 3D Radiomic Clinical dataset (balanced accuracy: 0.65 ± 0.19, AUC: 0.60 ± 0.20, 10-fold cross-validation). However, significant misclassification of fibrotic lesions was observed, likely due to dataset imbalance. The study's main limitation was the small sample size, impacting robustness and lesion characterization. These findings highlight the challenge of distinguishing active from fibrotic lesions in deep endometriosis using radiomic and clinical features. Future research should explore additional radiomic features and alternative modeling approaches to enhance classification performance and diagnostic potential.