Session Item

Urethra damage has been shown to be highly associated with urinary toxicity after prostate cancer radiotherapy. Identifying the urethra is therefore essential for efficient treatment planning, but is still challenging as this structure is completely invisible in the CT scans. Although atlas-based approaches have been proposed, their long execution time limits their clinical application. Here, we propose a Deep Learning (DL) based method for segmenting the intra-prostatic urethra from CT and compare it with a state-of-the-art model in terms of quality and time. The method exploits nearby OARs contours to anatomically guide the model, which learns the location of the urethra directly from computed meaningful Distance Maps (DMs).

Manual contours of the prostate, bladder and urinary catheters from 55 CT images of patients treated with brachytherapy and rigidly registered to a common space were randomly divided into training (80%) and testing (20%) datasets. The binary masks of the prostate and the bladder were employed to extract fixed volumes of interest (VOIs) and calculate DMs. For that, different combinations of the Euclidean distance transform and the Laplacian descriptor to the centroid of each organ were considered. These distance images were then used to train the so-called nnU-Net, a DL-based segmentation algorithm that includes automatic configuration and 5-fold cross-validation. To evaluate the concordance between the segmented urethra and the ground-truth catheter we measured Dice similarity coefficient (DSC), average surface distance (ASD), centerline distances (CLD) and the percentage of centerline lying in a certain radius of the catheter (PWR). The segmentation framework (Fig. 1) was applied to an independent database of 32 patients treated with IMRT at 78/80 Gy to quantify dose on the urethra.

The test performance of our approaches is summarized in Fig. 2. For all the metrics, the nnU-Net shows lower mean error values and lower variability compared to an existing multi-atlas based strategy (MABUS), with a significant improvement of both the CLD (1.63mm) and PWR within a radius <3.5mm (41%) and <5mm (14%) for our best model using the Euclidean distance and a weighted approach. Each 3D CT image, including DM computation and inference time, took an average runtime of 1.8 mins on a Nvidia RTX 8000 GPU. In IMRT, the urethra received higher doses (mean D95%_urethra -dose to 95% of the urethra- was 77.6Gy) than the whole prostate (mean D95%_prostate=76.6Gy).

We propose a DL-based solution that provides accurate and fast segmentation of the urethra. Our results showed that DL takes advantage of DMs to predict invisible regions in the CT scans, such as the urethra. The inputs of the proposed framework are only segmentation masks, so the translation into the MR domain would be straightforward. This method can be potentially applied to the clinical workflow to devise personalized treatment with reduced urinary toxicity in prostate radiation therapy.