Towards Automated Dosimetric Analysis of the Bladder Trigone: Deep-Learning-Based Joint Segmentation and Landmark Localization

Abstract

The bladder trigone dosimetry is hypothesized to have a stronger correlation with post-SBRT urinary toxicity than that of the entire bladder. However, the trigone tends to move significantly between simulation and daily treatment. Its small size, large daily motion, and proximity to the target lead to potentially consequential but unaccounted-for dosimetric uncertainties. Manual segmentation of the structure can be inconsistent and time-consuming, even with MR-guided RT. Here, we propose and demonstrate a deep-learning-based framework for joint segmentation and landmark localization to support deformable registration and comprehensive dosimetric analysis. A total of 30 patients were randomly selected for training, and 20 were held out for testing. Each patient had 1 simulation and 5 daily pre-treatment images obtained from a clinical 0.35T MR Linac. The trigone is defined as the triangular bladder section among three landmarks (2 ureteral orifices and the internal urethral orifice). In the manual contouring process, the 3 landmarks were identified first, followed by trigone segmentation. The proposed joint method uses a modified 3D nnU-Net with 2 decoders, one for segmentation and the other for landmark localization. The shared encoder is expected to extract features useful for both tasks. The joint framework was compared with a baseline method using two separate 3D nnU-Nets for landmark localization and trigone segmentation, respectively. Since the trigone is small (∼2% of the bladder volume), we further experimented with a second-stage prediction mimicking the human contouring process. The predicted landmarks from the first stage were used to crop the trigone region, and a second network was trained on cropped images. Evaluation metrics included the Dice score, 95% Hausdorff distance (HD95), and average surface distance (ASD) for segmentation, and Euclidean distance (ED) between the predicted and ground truth landmarks for localization. The quantification metrics are summarized in the table below. The joint approach shows similar Dice performance to the baseline method but markedly better HD95 by 13%. For landmark localization, the proposed method is similar to the baseline, but the integration of the segmentation task stabilizes the training process. The two-stage approach further improves HD95, ASD, and ED by 27%, 24%, and 19%. Combining segmentation and landmark localization exhibits a synergistic effect. The proposed two-stage approach provided additional improvement. Future studies will explore the deformable registration of the trigone based on the segmentation and landmark detection, as well as analyze cumulated dose distribution.