Diffusion Histology Imaging to Improve Lesion Detection and Classification in Multiple Sclerosis

Background: Diagnosing MS through magnetic resonance imaging (MRI) requires extensive clinical experience and tedious work. Furthermore, MRI-indicated MS lesion locations rarely align with the patients9 symptoms and often contradict with pathology studies. Our lab has developed and modified a novel diffusion basis spectrum imaging (DBSI) technique to address the shortcomings of MRI-based MS diagnoses. Although primary DBSI metrics have been demonstrated to be associated with axonal injury/loss, demyelination and inflammation, a more detailed analysis using multiple DBSI-structural metrics to improve the accuracy of MS lesion detection and differentiation is still needed. Here we report that Diffusion Histology Imaging (DHI), an improved approach that combines a deep neural network (DNN) algorithm with improved DBSI analyses, accurately detected and classified various MS lesion types. Methods: Thirty-eight multiple sclerosis patients were scanned with T2-weighted imaging (T2WI) using fluid attenuated inversion recovery (FLAIR), T1-weighted imaging (T1WI) using magnetization-prepared rapid acquisition with gradient echo (MPRAGE), magnetization transfer contrast (MTR) imaging and diffusion-weighted imaging. The imaging results identified 43,261 voxels from 91 persistent black hole (PBH) lesions, 89 persistent gray hole (PGH) lesions, 16 acute gray hole (AGH) lesions, 189 non-black hole (NBH) lesions and 113 normal-appearing white matter (NAWM) areas. Data extracted from these lesions were randomly split into training, validation, and testing groups with an 8:1:1 ratio. The DNN was constructed with 10 fully-connected hidden layers using TensorFlow 2.0 in Python. Batch normalization and dropout regularization were used for model optimization. Results: Each MS lesion type had unique DBSI derived diffusion metric profiles. Based on these DBSI diffusion metric profiles, DHI achieved a 93.6% overall concordance with neurologist determinations of all five MS lesions, compared with 74.3% from conventional MRI (cMRI)-DNN model, 78.2% from MTR-DNN model, and 80% from DTI-DNN model. DHI also achieved greater performances on detecting individual MS lesion types compared to other models. Specifically, DHI showed great performances on prediction of PBH (AUC: 0.991; F1-score: 0.923), PGH (AUC: 0.977; F1-score: 0.823) and AGH (AUC: 0.987; F1-score: 0.887), which significantly outperformed other models. Conclusions: DHI significantly improves the detection and classification accuracy for various MS lesion types, which could greatly aid the clinical decisions of neurologists and neuroradiologists. The efficacy and efficiency of this DNN model shows great promise for clinical application.