Exploring multi-b-value diffusion-weighted imaging for more accurate prostate segmentation and uncertainty estimation

Abstract

Background and objectiveAccurate auto-segmentation on diffusion-weighted imaging (DWI) remains a challenge. A 3D convolutional neural network (CNN) model using DWI was developed to automatically segment the prostate gland and estimate uncertainty in precision treatment planning. MethodsA total of 303 patients who uncervent multi-b-value DWI were retrospectively recruited and divided into training and testing groups based on the examination time. Radiologists labeled the prostate region on DWI images of b = 800 sec/mm2. A 3D U-shaped CNN was trained on multi-b-value DWI images (multi-b-value model) and evaluated using the Dice similarity coefficient (DSC), the Hausdorff distance (HD), and the average surface distance (ASD). Segmentation results were compared to a single-b-value model and registration result from T2-weighted images. The study estimated and analyzed the combination of epistemic and aleatoric uncertainty. ResultsThe DSC, HD, and ASD of our final prediction were 0.928 ± 0.036, 5.472 ± 4.008 mm, and 0.575 ± 0.508 mm. The proposed multi-b-value model achieved a higher DSC on DWI images with all b-values when comparing with the single-b-value model, and superior to the registration result on most of the b-values. The uncertainty helps to effectively filter out prediction errors and indicates that the model is more confident in the middle of the prostate gland, followed by the base and apex of the prostate gland. ConclusionsThe complementary information provided by the multi-b-value DWI images helps to improve both the performance of prostate segmentation, while also providing an estimation of the uncertainty associated with the segmentation prediction, which is clinically useful.