Deep Neural Network for Enhanced Intravoxel Incoherent Motion Fitting in Rectal Cancer

Motivation: In diffusion MRI of rectal cancer, characterized by inherently low SNR further compromised by intestinal peristalsis, the issue of poor precision and reproducibility in IVIM fitting needs to be addressed. Goal(s): Our goal was to improve the precision and reproducibility of IVIM fitting in rectal cancer. Approach: Diffusion MRI data from 116 rectal cancer patients were used to train deep neural network (DNN) for IVIM fitting, with comparisons made to conventional least‐squares regression (LSR) and Bayesian algorithms. Results: The DNN improved IVIM parameters maps quality, interobserver reproducibility, and provided better differentiation among different tissue types and clinical stages. Impact: The DNN demonstrated superior performance compared to LSR and Bayesian algorithms, significantly enhancing the precision and reproducibility of the IVIM fitting, thereby potentially improving the clinical applicability of IVIM model in rectal cancer.