A Deep Learning Approach for Automated Segmentation of Kidneys and Exophytic Cysts in Individuals with Autosomal Dominant Polycystic Kidney Disease

Abstract

BACKGROUND: Total kidney volume (TKV) is an important imaging biomarker in autosomal dominant polycystic kidney disease (ADPKD). Manual computation of TKV, particularly with the exclusion of exophytic cysts, is laborious and time consuming. METHODS: We developed a fully automated segmentation method for TKV using a deep learning network to selectively segment kidney regions while excluding exophytic cysts. We used abdominal T2 -weighted magnetic resonance images from 210 individuals with ADPKD who were divided into two groups: one group of 157 to train the network and a second group of 53 to test it. With a 3D U-Net architecture using dataset fingerprints, the network was trained by K-fold cross-validation, in that 80% of 157 cases were for training and the remaining 20% were for validation. We used Dice similarity coefficient, intraclass correlation coefficient, and Bland-Altman analysis to assess the performance of the automated segmentation method compared with the manual method. RESULTS: The automated and manual reference methods exhibited excellent geometric concordance (Dice similarity coefficient: mean+/-SD, 0.962+/-0.018) on the test datasets, with kidney volumes ranging from 178.9 to 2776.0 ml (mean+/-SD, 1058.5+/-706.8 ml) and exophytic cysts ranging from 113.4 to 2497.6 ml (mean+/-SD, 549.0+/-559.1 ml). The intraclass correlation coefficient was 0.9994 (95% confidence interval, 0.9991 to 0.9996; P<0.001) with a minimum bias of -2.424 ml (95% limits of agreement, -49.80 to 44.95). CONCLUSIONS: We developed a fully automated segmentation method to measure TKV that excludes exophytic cysts and has an accuracy similar to that of a human expert. This technique may be useful in clinical studies that require automated computation of TKV to evaluate progression of ADPKD and response to treatment.