Deep-learned estimation of uncertainty in measurements of apparent diffusion coefficient from whole-body diffusion-weighted MRI.

Abstract

PURPOSE: To use deep learning to calculate the uncertainty in apparent diffusion coefficient (σADC) voxel-wise measurements to clinically impact the monitoring of treatment response and improve the quality of ADC maps. MATERIALS AND METHODS: We use a uniquely designed diffusion-weighted imaging (DWI) acquisition protocol that provides gold-standard measurements of σADC to train a deep learning model on two separate cohorts: 16 patients with prostate cancer and 28 patients with mesothelioma. Our network was trained with a novel cost function, which incorporates a perception metric and a b-value regularisation term, on ADC maps calculated by combinations of 2 or 3 b-values (e.g. 50/600/900, 50/900, 50/600, 600/900 s/mm2). We compare the accuracy of the deep-learning based approach for estimation of σADC with gold-standard measurements. RESULTS: The model accurately predicted the σADC for every b-value combination in both cohorts. Mean values of σADC within areas of active disease deviated from those measured by the gold-standard by 4.3% (range, 2.87-6.13%) for the prostate and 3.7% (range, 3.06-4.54%) for the mesothelioma cohort. We also showed that the model can easily be adapted for a different DWI protocol and field-of-view with only a few images (as little as a single patient) using transfer learning. CONCLUSION: Deep learning produces maps of σADC from standard clinical diffusion-weighted images (DWI) when 2 or more b-values are available.