A deep learning approach to evaluate intestinal fibrosis in magnetic resonance imaging models

Fibrosis may be introduced as a severe complication of inflammatory bowel disease (IBD). This is a particular disorder causing luminal narrowing and stricture formation in the inflamed bowel wall of a patient denoting, possibly, need for surgery. Thus, the development of treatments reducing fibrosis is an urgent issue to be addressed in IBD. In this context, we require the finding and development of biomarkers of intestinal fibrosis. Potential candidates such as microRNAs, gene variants or fibrocytes have shown controversial results on heterogeneous sets of IBD patients. Magnetic resonance imaging (MRI) has been already successfully proven in the recognition of fibrosis. Nevertheless, while there are no numerical models capable of systematically reproducing experiments, the usage of MRI could not be considered a standard in the inflammatory domain. Hence, there is an importance of deploying new sequence combinations in MRI methods that enable learning reproducible models. In this work, we provide reproducible deep learning models of intestinal fibrosis severity scores based on MRI novel radiation-induced rat model of colitis that incorporates some unexplored sequences such as the flow-sensitive alternating inversion recovery or diffusion imaging. The results obtained return an 87.5% of success in the prediction of MRI scores with an associated mean-square error of 0.12. This approach offers practitioners a valuable tool to evaluate antifibrotic treatments under development and to extrapolate such noninvasive MRI scores model to patients with the aim of identifying early stages of fibrosis improving patients' management.