Motivation: Conventional IVIM fitting methods are susceptible to noise and potentially unreliable in clinical decision-making. Recent work showed promise for spatially-aware deep learning. However, its clinical efficacy remains unexplored. Goal(s): This research aims to evaluate whether transformer-based networks can provide superior IVIM parameter estimation, specifically improving tumor classification accuracy across tumor grades. Approach: We compared conventional estimators and advanced transformer-based algorithms trained on synthetic data. Performance was assessed in both simulations and a brain tumor cohort, focusing on tumor classification accuracy. Results: The transformers demonstrated higher accuracy in simulations, provided superior tumor delineation in pseudo-diffusion parameter maps, and improved classification accuracy of high-grade tumors. Impact: This study shows that transformer-based model fitting offers clinically valuable IVIM parameter estimates and potential for enhanced tumor classification accuracy. This advancement improves the noninvasive assessment of tumor heterogeneity, bringing IVIM closer to clinical use and supporting personalized treatment strategies.
Kaandorp et al. (Tue,) studied this question.