Motivation: Non-linear least-squares (NLLS) fitting to quantify DCE-MRI is sensitive to noise, non-convex, and slow, leading to unreliable parameter estimation especially for low-dose/low-SNR protocols. Goal(s): To enhance the reliability and efficiency of pharmacokinetic parameter fitting in DCE-MRI using AI. Approach: We trained a deep neural network using extensive simulations across 20%- to full-dose scenarios and evaluated its performance in in-vivo multitasking DCE-MRI data. Results: AI-based fitting provides pharmacokinetic parameters that (1) align with literature values, (2) improve tumor vs. healthy pancreas differentiation, (3) offer greater homogeneity in healthy pancreas regions, and (4) reduce processing time. Impact: Deep learning significantly improves the precision, homogeneity, and speed of pharmacokinetic fitting in DCE-MRI, making it an attractive alternative to NLLS. This advancement supports more efficient, accurate, and clinically feasible quantitative imaging for various biomedical applications.
Wu et al. (Tue,) studied this question.