Sequence-robust MRI asymmetry measures for the detection of anoperineal lesions in Crohn's disease on axial pelvic MRI scans.

Magnetic resonance imaging (MRI) is widely considered the gold standard for evaluating Crohn's disease anoperineal lesions. However, despite its central role in clinical evaluation, automated methods specifically dedicated to the analysis of these lesions remain rare. This is particularly challenging in the context of anoperineal lesions due to (i) the diversity of MRI sequences, (ii) the small size of cohorts available to train complex models, and (iii) the absence of a standard shape for lesions. In this context, we explore an alternative strategy based on interpretable asymmetry features in axial MRIs, explicitly capturing morphological and intensity differences between the left and right sides of the perineal region. These features were used to either classify images as containing or not anoperineal lesions, or to precisely locate a lesion on an MRI slice. We apply this method on a small cohort of 44 patients affected by Crohn's disease, including 134 sequences of MRI of various types (38 T1 sequences, 74 T2 sequences, etc). First, a proof-of-concept dataset was constructed, including images either clearly presenting lesions or not (thus excluding symmetrical lesions), to prove that the asymmetry features we compute are sensitive to the presence of asymmetric lesions on MRI images. Importantly, we were able to show that the asymmetry features are not particularly sensitive to MRI sequences' type, or patients' effect, even with limited annotated datasets. The classification model demonstrates promising diagnostic performance, achieving an Area Under Curve (AUC) greater than 0.85 on the proof-of-concept dataset. However, when making predictions on the validation and test dataset, the AUCs drop to 0.667 and 0.647 respectively. The model struggles to generalize to all kinds of lesions, particularly to classify lesions slices when symmetrical (horseshoe) lesions are observed. Additionally, the method designed to locate a lesion in an image exhibited good precision, with a median distance of 1 centimeter between the detected area and the actual lesion position on our proof-of-concept dataset and 1.2 centimeter on the complete dataset, indicating its potential usefulness in future studies focused on standardized assessment of anoperineal lesions.