Kermanshah University of Medical Sciences
Accurate delineation of prostate and organs at risk (OARs) in ultrasound images plays an important role in HDR brachytherapy treatment planning but due to the low soft contrast tissue, it is very time-consuming and challenging task and prone to inter and intra-observer variations. In this work, we propose a two-stage deep learning-based approach for fast, reliable, and reproducible auto-segmentation of prostate and OARs in HDR brachytherapy.
In this work, we developed a Cascade-Net which is consists of two states of art neural networks for the prostate, urethra, and rectum segmentation in ultrasound images. In our segmentation framework, the first stage is the organ localizer module, which generates a candidate segmentation region of interest (ROIs) for each organ. The second stage produces a more robust and accurate contour from the previous coarse segmentation mask. A U-Net with an attention mechanism on skip connections and a deep supervision concept will generate ROIs by eliminating irrelevant background information. This network will identify the probability of the presence of each organ. The extracted regions will be fed to the attention deeplab3 to generate a fine segmentation. Ultrasound Images of 109 patients with prostate cancer were utilized in this study. The performance of the proposed framework was evaluated through well-established quantitative metrics such as Dice similarity coefficient (DSC), and Hausdorff distance (HD).
The Cascade-Net framework achieved the segmentation results with a DSC of 96 ± 3%, 94 ± 1%, 92 ± 2 for prostate, Urethra, and rectum, respectively. The HD values (mm) were 0.04 ± 0.01, 0.03 ± 0.02, 0.04 ± 0.03 for prostate, Urethra, rectum, respectively. There was no statistically significant difference between manual segmentation and the Cascade-Net framework (P-value > 0.05).
The results of our study demonstrate that our auto-contouring segmentation framework can be used for fast, reliable, and reproducible segmentation of the prostate and OARs to facilitate the brachytherapy workflow.