Development and acceptability validation of a deep learning-based tool for whole-prostate segmentation on multiparametric MRI: a multicenter study

Background: Accurate whole prostate segmentation on magnetic resonance imaging (MRI) is important in the management of prostatic diseases. In this multicenter study, we aimed to develop and evaluate a clinically applicable deep learning-based tool for automatic whole prostate segmentation on T2-weighted imaging (T2WI) and diffusion-weighted imaging (DWI). Methods: In this retrospective study, 3-dimensional (3D) U-Net-based models in the segmentation tool were trained with 223 patients who underwent prostate MRI and subsequent biopsy from 1 hospital and validated in 1 internal testing cohort (n=95) and 3 external testing cohorts: PROSTATEx Challenge for T2WI and DWI (n=141), Tongji Hospital (n= 30), and Beijing Hospital for T2WI (n=29). Patients from the latter 2 centers were diagnosed with advanced prostate cancer. The DWI model was further finetuned to compensate for the scanner variety in external testing. A quantitative evaluation, including Dice similarity coefficients (DSCs), 95% Hausdorff distance (95HD), and average boundary distance (ABD), and a qualitative analysis were used to evaluate the clinical usefulness. Results: The segmentation tool showed good performance in the testing cohorts on T2WI (DSC: 0.922 for internal testing and 0.897-0.947 for external testing) and DWI (DSC: 0.914 for internal testing and 0.815 for external testing with fine-tuning). The fine-tuning process significantly improved the DWI model's performance in the external testing dataset (DSC: 0.275 vs. 0.815; P<0.01). Across all testing cohorts, the 95HD was <8 mm, and the ABD was <3 mm. The DSCs in the prostate midgland (T2WI: 0.949-0.976; DWI: 0.843-0.942) were significantly higher than those in the apex (T2WI: 0.833-0.926; DWI: 0.755-0.821) and base (T2WI: 0.851-0.922; DWI: 0.810-0.929) (all P values <0.01). The qualitative analysis showed that 98.6% of T2WI and 72.3% of DWI autosegmentation results in the external testing cohort were clinically acceptable. Conclusions: The 3D U-Net-based segmentation tool can automatically segment the prostate on T2WI with good and robust performance, especially in the prostate midgland. Segmentation on DWI was feasible, but fine-tuning might be needed for different scanners.

Automated summary

In this study, a deep learning-based tool was developed to automatically segment the whole prostate on T2WI and DWI. The tool showed good performance on T2WI and feasible results on DWI with fine-tuning. It had better segmentation results in the prostate midgland.