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

Deep learning (DL)-based approaches have recently become powerful tools for characterizing microstructural properties of brain tissue from diffusion-weighted magnetic resonance imaging data. These techniques enable depicting microstructural features with significantly reduced acquisition schemes while preserving relatively high image-quality metrics compared with non-DL-based approaches. Learning-based methods have increasingly been applied to diffusion tensor imaging (DTI) and more advanced multi-compartmental biophysical models, with the potential to facilitate multiparameter analyses of brain microstructure in clinical settings. However, independently trained DL models-while maintaining image-quality properties-do not appear to preserve the statistical properties of groupwise analyses across independently trained models. In this study, we systematically evaluate the statistical properties of DL-based DTI parameters across independently trained neural networks under the same architecture to preserve group-level effects in bipolar disorder. We examine the impact of model retraining and its dependence on reducing the number of gradient directions compared with the nonlinear least-squares (NLLS) procedure. Our results show that initializing model weights randomly from a uniform distribution, while keeping the network architecture and training hyperparameters the identical in the experiment, leads to significant variation in the statistical properties of DTI measures inferred from independently trained models. The analyses showed that the true positive rate in groupwise analysis using the DL-based DTI parameters ranges from 0\% to 100\% compared with the NLLS-based analysis. As the number of gradient directions increased, the NLLS procedure yielded more stable groupwise statistical results, whereas independently trained DL models showed greater variability despite similar image-quality agreement with reference maps. The results indicate that the DTI maps retrieved from a single trained network should be treated with great caution before drawing clinical conclusions.