Topological Radiomics in Predicting Mucosal Healing and Activity Score of Crohn’s Disease Using Multi-Task Deep Learning

Kaiqiang Tang,1,* Xinyan Wu,1,2,* Junlin Li,3 Li Yu,1,4 Ling Yang,1 Yue Zhang,1,4 Lingfeng Zhang,1,2 Yazhou Wang,5 Kang Li1,4 1Department of Radiology, Chongqing General Hospital, Chongqing University, Chongqing, 401147, People’s Republic of China; 2North Sichuan Medical College, Nanchong, Sichuan, 637100, People’s Republic of China; 3Department of Radiology, The 13th People’s Hospital of Chongqing, Chongqing, 400050, People’s Republic of China; 4Chongqing Medical University, Chongqing, 400016, People’s Republic of China; 5Chongqing University, Chongqing, 400044, People’s Republic of China*These authors contributed equally to this workCorrespondence: Kang Li, Department of Radiology, Chongqing General Hospital, Chongqing University, Chongqing, 401147, People’s Republic of China, Email [email protected]: Endoscopic evaluation remains the gold standard for assessing Crohn’s disease (CD) activity and mucosal healing (MH), but it is invasive, expensive and time consuming. Therefore, there is an urgent need for a non-invasive quantitative alternative method.Aim: To develop a topological radiomics-based multi-task deep learning model for simultaneous prediction of MH status and endoscopic activity scores in CD.Methods: A total of 81 CD patients were stratified into training (n=60) and validation (n=21) groups at a 7:3 ratio. Topological radiomic features were extracted from multiphase CT enterography. A multi-task model was trained to predict MH (classification) and SES-CD (regression), integrating feature selection and SHAP-based interpretability.Results: Three discriminative topological features were identified across arterial and portal phases. For MH prediction, the multi-task model achieved an AUC of 0.938 for training set and 0.875 for validation set. For SES-CD prediction, it showed lower MSE and MAE, with higher R2 and C-index than the single-phase models.Conclusion: The multi-task topological radiomics framework enables accurate, non-invasive assessment of mucosal healing and endoscopic activity in CD, offering a clinically interpretable approach with strong translational potential. Future studies with larger cohorts are warranted to further validate its robustness.Keywords: topological data analysis, Crohn’s disease, deep learning, mucosal healing