Evaluation of Generative Deep Learning for Synthesizing High B-Value Diffusion-Weighted MRI of Musculoskeletal Tumors
Steven Zhang, Sara Ataie, Emily Squires, Shanesh Kumar, Kerolos Hendy, Alice Andrawos, Wan Yin Lim, Saba Hatami, Andrew Dwyer, Angela Walls, Ganessan Kichenadasse, Jakub Jagiello, Luke Johnson, Vu Minh Hieu Phan, Minh-Son ToAbstract
Background
In preoperative magnetic resonance imaging assessment of bone and soft tissue tumors, diffusion-weighted imaging (DWI) estimates tumor cellularity and aggressiveness. High b-value DWI provides greater tumor conspicuity, but is not widely available.
Purpose
Explore the feasibility of deep learning for synthesizing high b-value DWI and apparent diffusion coefficient (ADC) for pre-operative evaluation of bone and soft tissue tumors.
Materials and Methods
30 adults planned for surgery or biopsy were recruited prospectively between January 15, 2024 and August 2, 2024 at a single center. Deep learning models were trained separately for high b-value DWI and ADC synthesis from low b-value DWI and T2-weighted images using a patient-level train-validation-test split design. Synthesis quality compared to ground truth quantitatively assessed using performance metrics included, learned perceptual image patch similarity (LPIPS), structural similarity index measure (SSIM), peak signal noise ratio (pSNR) and mean squared error (MSE). Synthesis quality was also assessed by four board-certified radiologists via a 5-point Likert scale and pairwise comparisons. Statistical analysis involved Wilcoxon Signed-Rank test for synthesis quality and a Friedman test with post hoc Nemenyi test for model ranking.
Results
The top 3 performing models were conditional Stochastic Differential Equations (cSDE), Pix2Pix Turbo, and Discriminator Edge-Aware Generative Adversarial Network (dEa-GAN). Reconstruction metrics (SSIM, pSNR, LPIPS, MSE) were comparable across models for both synthesized DWI and ADC maps. Likert ratings from four radiologists had a mode of 4 and median of 4 (DWI: p=0.00318, p<0.001; ADC: p<0.001, p=0.00348) for cSDE and dEa-GAN, indicating favorable image quality, with pairwise comparisons showing preference for cSDE.
Conclusions
In this feasibility study, deep learning showed promise for synthesizing high b-value DWI and ADC and should be further developed with validation on larger external datasets and task-based evaluations.