Clinical relevance of diffusion tensor imaging-based parameters estimated with deep learning: the effect of model retraining

Randomly initialized deep learning models estimating diffusion tensor imaging parameters in bipolar disorder yield groupwise statistical properties that vary widely, with true positive rates ranging from 0% to 100% versus NLLS.

• Why it matters: Independently trained DL models may preserve image-quality metrics yet fail to maintain consistent group-level statistics, undermining clinical interpretation in bipolar disorder.
• What they did: The study evaluated statistical properties of DL-based DTI parameters across independently trained neural networks with the same architecture, testing how model retraining and gradient-direction reduction affect results compared with nonlinear least-squares (NLLS).
• The result: Increasing gradient directions made NLLS groupwise results more stable, while independently trained DL models remained more variable despite similar agreement with reference maps, implying single-network DTI maps need caution for clinical conclusions.

AI summary of the abstract · may contain errors · not medical advice · read the paper.