Abstract Background Radiomics applies machine learning to extract quantitative imaging features beyond the limits of visual assessment. We aimed to combine radiomics and clinical data to predict the likelihood of surgical intervention in patients with Crohn’s disease (CD). Methods We utilized MRE scans and clinical data from the epi-IIRN cohort that includes all patients with inflammatory bowel disease (IBD) in Israel (N = 74,372). The model was first trained on 107 annotated MRE scans from a public dataset of patients with CD. Bowel regions were automatically identified using a nnU-Netv2 model (Figure). We included patients from the epi-IIRN who had an MRE scan around the time of diagnosis. Outcome was defined as the need for surgical intervention. From each bowel region, 101 radiomics features describing intensity, texture, and shape were extracted. An architecture based on a Graph Neural Network–Cox (GNN-Cox) model was selected representing the bowel as a spatially connected structure with model interactions between neighboring segments. Radiomics features were combined with clinical information. Results A total of 492 patients were included (241 (48%) pediatric-onset, 251 (52%) adult-onset). Follow-up time was censored at 70 months, with a median follow-up of 42 months, during when 40 patients (8%) required surgical intervention. The segmentation algorithm showed higher Dice score (reflects the overlap between the predicted region and the manual annotation) for bowel regions with more pixels (e.g. Jejunum and ileum; 0.77 and 0.82), and lower scores for regions like the rectum (0.51). The GNN–Cox model achieved a C-index of 0.66 (95%CI 0.59–0.73), with an AUC of 0.70 (95%CI 0.52–0.78) at 12 months and 0.69 (95%CI 0.58–0.77) at 36 months. In comparison, using the same input parameters, a tree-based baseline model (Gradient Boosting Survival Analysis) achieved an AUC of 0.55 (95%CI 0.46-0.63) at 12 months. In the final model, radiomic features from the ileum and jejunum emerged as the most informative contributors to the predictive graph. Conclusion We demonstrate the potential of radiomics for predicting bowel resection in Crohn’s disease. Incorporating spatial relationships between bowel segments improved performance compared with other models. Most predictive signal raised from the ileum and jejunum. Conflict of interest: Kurtser, Gili: No conflict of interest Amit, Yosef: No conflicts of interests Badash, Zvi Natan: No conflict of interest Vainberg, Elena: Hagopian, Talar: No conflict of interest Gotesdyner, Leora: No conflict of interest Yogev, Dotan: No conflict of interest Focht, Gili: Other: Received in the past 3 years consultation fee from Lilly Ben-Tov, Amir: No conflict of interest Zacay, Galia: No conflict of interest Matz, Eran: No conflict of interest Zekaria, Shir: No conflict of interest Bar-Yoseph, Haggai: No conflict of interest Gavrielov, Naama: No Gitelman, Leah: No conflict of interest Dotan, Iris: Grant: The Leona M. and Harry B. Helmsley Charitable Trust, Altman Research, Pfizer, BMS Personal Fees: Pfizer, Falk, Ferring, Abbvie, Janssen, Celltrion, Takeda, Celgene/BMS, Gilead, Galapagos, Materia Prima, Sandoz, Sublimity, Sangamo, Spyre, Eli-Lilly, Harp Diagnostics, Gutreat, Astra Zeneca Cytter-Kuint, Ruth: No conflict of interest Freiman, Moti: No conflict of interest Turner, Dan: Consultation fee: Janssen, Pfizer, Ferring, Abbvie, Takeda, Prometheus Biosciences, Lilly, SorrisoPharma, Boehringer Ingelheim, Galapagos, BMS, AlfaSigma, Merck, Gentech Research support: Janssen, Abbvie, Takeda, Pfizer Royalties: Shaare Zedek Medical Center, Hospital for Sick Children
Kurtser et al. (Thu,) studied this question.