AI-Enhanced MRI Radiomics for Discriminating Active and Fibrotic Lesions to Support Therapeutic Decision-Making in Deep Endometriosis.
Authors
Affiliations (6)
Affiliations (6)
- Department of Medical and Surgical Sciences, University of Bologna, 40138 Bologna Italy; IRCCS Azienda Ospedaliero Universitaria di Bologna, 40138 Bologna Italy.
- Department of Medical and Surgical Sciences, University of Bologna, 40138 Bologna Italy; Department of Radiology, IRCCS Azienda Ospedaliero-Universitaria di Bologna, 40138 Bologna Italy.
- Department of Medical and Surgical Sciences, University of Bologna, 40138 Bologna Italy; Division of Gynecology and Human Reproduction Physiopathology, IRCCS Azienda Ospedaliero-Universitaria di Bologna, 40127 Bologna Italy.
- Department of Physics and Astronomy, University of Bologna, 40127 Bologna Italy; INFN Bologna Italy. Electronic address: [email protected].
- IRCCS Istituto delle Scienze Neurologiche di Bologna, Data Science and Bioinformatics Laboratory, 40139 Bologna Italy.
- Department of Woman, Child, and General and Specialized Surgery, University of Campania "Luigi Vanvitelli", 80138 Naples, Italy.
Abstract
This study investigates whether quantitative analysis of preoperative Magnetic Resonance Imaging (MRI) scans can differentiate deep infiltrating endometriosis (DIE) lesion types (active or fibrotic) and associate them with reported pain symptoms. monocenter, observational, retrospective accuracy study SETTING: Academic Hospital IRCCS Policlinico di S. Orsola Bologna, Italy PARTICIPANTS: All patients with histologically confirmed DIE who had undergone pre-operative MRI were considered for recruitment during the study period (12/2022-12/2023). Histological assessment was performed on all surgical specimens categorizing the lesions in active lesions (AL) and fibrotic lesions (FL). All patients underwent pelvic MRI prior to surgery, and DICOM files were retrieved. Radiomic features were extracted using PyRadiomics (v3.1.0). In addition to radiomic data, morphological features were extracted. A genetic algorithm (GA) was used for features selection to identify relevant features and build classification models. Machine learning (ML) pipelines, tuned using leave-one-out cross-validation and evaluated via balanced accuracy, Matthews Correlation Coefficient, sensitivity and specificity, were developed to identify the most informative morphological and textural features for lesion classification and pain severity classification. Sixty-one women were enrolled. The integration of radiomic and graphomic enabled the extraction of insightful imaging features statistically linked to histological lesion types and pain symptomatology. The ML models proved promising classification performance under the leave-one-out cross-validation framework, identifying imaging biomarkers that differentiate fibrotic lesions from active stromal ones (balanced accuracy: 72.7%) and correlate with pain profiles (reaching a maximum of 88.0% of balanced accuracy for chronic pelvic pain). These results highlight the potential of imaging-based phenotyping in DIE. The results support the feasibility of a non-invasive, image-based approach to preoperatively characterize DIE lesions. By combining radiomic, graphomic, and ML, our results showed that a rich characterization of active or fibrotic DIE lesions could enable their differentiation and their association with pain symptoms.