Geometric evaluation of a deep learning method for segmentation of urinary OARs on magnetic resonance imaging for prostate cancer radiotherapy.

INTRODUCTION

While urinary organs at risk (OARs) such as the intraprostatic urethra and the bladder trigone are increasingly recognized as associated with severe genitourinary toxicity, their delineation in clinical practice is time consuming and probably associated with a large interobserver variability. The aim of this study was to propose a magnetic resonance (MR) deep learning segmentation of urinary OARs for prostate cancer (PCa) radiotherapy (RT), based on a validated atlas.

MATERIAL AND

METHODS

In this multicentric study, a convolutional neural network (CNN) for image segmentation (nnU-Net) was trained and validated on three image datasets. Two datasets came from MR-linac devices (Unity®, Elekta and MRIdian®, Viewray), and one dataset came from the PROSTATEx database (MAGNETOM® Trio and Skyra, Siemens). Evaluation of the deep learning segmentation was performed using dice score coefficients (DSC), surface distance (SD) and Hausdorff distance.

RESULTS

A total of 265 MRI were analyzed. The mean DSC for all urinary structures was 0.88. The automatic segmentation model proved to be effective in the segmentation of the target volume and large OARs such as the bladder (mean DSC ranging of 0.95). Regarding urinary OARs, the mean DSC ranged between 0.50 and 0.68. The Hausdorff distance ranged between 4.0 mm to 10.3 mm for urinary OARs, highlighting local mismatches caused by large anatomical variations between patients. However, the SD ranged between 1.0 mm and 1.3 mm for urinary OARs, highlighting an overall good surface correlation for all organs.

CONCLUSION

This multicentric study is the first to propose a nnU-Net deep learning model for the delineation of urinary OARs, that can be applied to various image dataset. Further work is needed to assess the dosimetric impact of such variations, in various clinical scenarios.