Revisiting prostate segmentation in magnetic resonance imaging (MRI): On model transferability, degradation and PI-RADS adherence

Abstract

AbstractPurposeTo investigate the effect of scanner and prostate MRI acquisition characteristics when compared to PI-RADSv2.1 technical standards in the performance of a deep learning prostate segmentation model trained with data from one center (INST1), longitudinally evaluated at the same institution and when transferred to other institutions.Materials and MethodsIn this retrospective study, a nn-UNet for prostate MRI segmentation was trained with data from 204 patients from one institution (INST1) (0.50mm2in-plane, 3.6mm thickness and 16cm field of view [FOV]). Post-deployment performance at INST1 was tested with 30 patients acquired with a different protocol and in a different period of time (0.60mm2in-plane, 4.0mm thickness and 19cm FOV). Transferability was tested on 248 patient sequences from five institutions (INST2, INST3, INST4, INST5 and INST6) acquired with different scanners and with heterogeneous degrees of PI-RADS v2.1 technical adherence. Performance was assessed using Dice Score Coefficient, Hausdorff Distance, Absolute Boundary Distance and Relative Volume Difference.ResultsThe model presented a significant degradation for the whole gland (WG) in the presence of a change of acquisition protocol at INST1 (DSC:99.46±0.12% and 91.24±3.32%,P<.001; RVD:-0.006±0.127% and 8.10±8.16%,P<.001). The model had a significantly higher performance in centers adhering to PI-RADS v2.1 when compared to those that did not (DSC: 86.24±9.67% and 74.83±15.45%,P<.001; RVD: -6.50±18.46% and 1.64±29.12%,P=.003).ConclusionsAdherence to PI-RADSv2.1 technical standards benefits inter-institutional transferability of a deep learning prostate segmentation model. Post-deployment evaluations are critical to ensure model performance is maintained over time in the presence of protocol acquisition modifications.