PHSOR02 Presentation Time: 9:05 AM

Purpose

High dose rate (HDR) prostate brachytherapy (BT) procedure requires imaging to guide transperineal needle insertion, either with CT, MR, or ultrasound (US) imaging. US is occasionally favored for its streamlined workflow and when access to other imaging is limited. General anesthesia is often used throughout the procedure, thus minimizing overall planning time is crucial to mitigate potential complications and allow for better management of operating room time. In this study, we explore the accuracy and time-saving potential of AI-driven auto-reconstruction of transperineal needles in the context of US-guided prostate BT planning.

Materials and Methods

A total of 98 US BT cases from a single institution were used in this work. US images were acquired using a BK3000 US + E14CL4b endocavity biplane transducer and combined into 3D-US datasets using the Oncentra Prostate system from Elekta. Gray value histogram of each 3D-US image was normalized. The data was split into 3 groups: 50 for training and validation (training set), 11 to evaluate reconstruction accuracy (test set #1) and 37 to evaluate the AI tool in a clinical implementation (test set #2). A 3D-UNet machine learning network was used, using human-reconstructed needles during the BT procedure as the reference segmentation mask. Model training was performed using the PyTorch library version 2.0.1 on a NVIDIA Quadro RTX 6000 GPU using Dice loss and AdamW optimizer. A 10-fold cross-validation scheme was employed during training. Reconstruction accuracy for test set #1 was evaluated by having 4 medical physicists manually reconstructing needles on the 3D-US scan after treatments. Ground truth reference needle positions for each observer (AI included) were determined from the other 4 reconstructions using a weighted voting average inspired by the STAPLE algorithm. Reconstruction accuracy was evaluated by taking the root mean squared error from the center of each reconstructed needle to the center of the ground truth needle, on each image axial slice in which the needle was visible by both humans and AI. Interobserver variability was evaluated using one-way ANOVA and Tukey's HSD post-hoc test. The needle total reconstruction time for test set #2 was taken as the timestamp difference from scan acquisition to final modification of the plan before dose calculations. This value was compared to values of the 50 cases done before the clinical implementation of the AI-assisted tool using a two-sample z-test. For this phase, we also measured the true positive rate of needle reconstruction and the # of AI-reconstructed needles that were further adjusted by the human planner.

Results

A mean error of (0.47±0.31) mm was found between the AI-reconstructed and the ground truth needles in test set #1, with 95.2% of AI needle points falling below 1 mm from their human-made counterparts. One-way ANOVA showed statistical difference between observers (p < 0.01), but post-hoc analysis showed only one of the human observers was significantly different from the others including the AI (α = 0.05). In test set #2, the AI algorithm achieved a true positive reconstruction rate of 93.7% (i.e. an average of 1.02 needles was missed per scan). Of these AI-reconstructed needles, only 5.5% required manual corrections from the planner before needle tip adjustment (using the needle length exiting from the template). Total time required to perform AI-assisted catheter reconstruction on clinical cases was on average 20.6 min, a decrease of 13.8 min (p < 0.01) compared to manual needle reconstruction as performed before the AI tool introduction.

Conclusions

This study demonstrates the feasibility and performance of an AI tool for transperineal needles reconstruction during 3D-US based HDR prostate BT. AI-generated catheters were within interobserver variability for all but one physicist. This methodology is a step toward treatment planning automation and increased efficiency in BT procedures.