Motivation: IVIM imaging is essential for distinguishing breast tumor subtypes, but traditional methods are noise-sensitive, and current deep learning approaches are limited by tissue heterogeneity. Goal(s): To propose a tissue-specific deep learning approach to optimize estimation performance compared to single-model estimation approach. Approach: Two stages: tissue segmentation and tissue-specific parameter estimation. U-Net model segments tissue types from DWI data. Next, MLPs are trained with masks for each tissuetype to estimate IVIM parameters, which are then combined for the final estimation. Results: With the segmentation result, the estimation results show better performance, and the distribution more closely aligned with the ground truth, particularly for tumors. Impact: IVIM model has gained momentum recently, especially in the field oncology. Our study improve the parameter estimation performance for IVIM model, which is important for tumor and tumor type prediction.
Liu et al. (Tue,) studied this question.