A Comparative Study of IVIM-MRI Fitting Techniques in Glioma Grading: Conventional, Bayesian, and Voxel-Wise and Spatially-Aware Deep Learning Approaches.

backgroundIntravoxel incoherent motion (IVIM) analysis of diffusion-weighted MRI (DWI) provides microvascular perfusion and diffusion information. However, parameter estimation is limited by noise sensitivity, variability across fitting methods, and lack of standardization. Deep-learning (DL)-based approaches, particularly spatially-aware transformers, may improve robustness, but their clinical utility remains unexplored. PURPOSE: To evaluate conventional, Bayesian, and DL-based IVIM fitting methods in glioma patients, focusing on tumor grading accuracy. STUDY TYPE: Retrospective. POPULATION: Fractal-noise-based simulations and preoperative DWI from 20 glioma patients (5 Grade-2, 3 Grade-3, 12 Grade-4). SEQUENCE: Spin-echo echo-planar DWI, 16 b values (0-900 s/mm ASSESSMENT: IVIM parameter maps were compared across least squares (LSQ), segmented (SEG), Bayesian shrinkage (BSP), and spatial-homogeneity (FBM) priors, and DL-based methods, including IVIM-NET, novel spatially-aware transformers (NATTEN-17), and a refined version (SA-17). Simulation accuracy was evaluated using median absolute percentage error (MDAPE) and bias using median percentage error. In vivo data were visually assessed by the authors for noise suppression and structural preservation. Whole-tumor diffusion coefficient (D), pseudo-diffusion coefficient (D*), and signal fraction (f) values were evaluated across tumor grades and for differentiating Grade-4 from Grade-2/3 tumors. STATISTICAL TESTS: Mann-Whitney U tests for group comparisons; tumor grading performance using receiver operating characteristic-area under the curve (ROC-AUC), and pairwise AUC differences using the DeLong test. SIGNIFICANCE: p < 0.05.

resultsTransformer-based methods achieved superior simulation accuracy, with significantly lower MDAPE for f and D* than all other approaches: NATTEN-17 (5.91%, 13.31%), SA-17 (7.73%, 13.66%), LSQ (21.95%, 54.34%), SEG (17.10%, 21.27%), BSP (12.35%, 22.79%), FBM (16.32%, 20.67%). In vivo, they provided superior visual quality and tumor delineation in f and D* maps, producing seemingly denoised versions of LSQ, while preserving tumor heterogeneity. The spatially-aware transformers yielded consistently the highest ROC-AUC values, particularly for D* (SA-17: 0.78), significantly outperforming LSQ (0.62), SEG (0.58), FBM (0.62), and IVIM-NET (0.71). DATA

conclusionTransformer-based model fitting has the potential to provide clinically valuable IVIM parameter estimates and improved tumor grading accuracy. EVIDENCE LEVEL: 3. TECHNICAL EFFICACY: Stage 2.