Implementation and Analysis of Deep Learning Needle Reconstruction Tool (CorneliUS) for US-Guided HDR Prostate Brachytherapy
Abstract
Purpose
To validate and assess the performance of an AI-assisted tool, CorneliUS, for needle reconstruction in high dose rate (HDR) ultrasound-guided prostate brachytherapy in an independent clinical setting.
Methods
CorneliUS’s performance was evaluated on a cohort of 49 HDR whole gland 15 Gy prostate brachytherapy patients from our institution. The ultrasound imaging equipment used was identical that of the original institution (BK Medical’s bk3000 US system, E14CL4b endocavity biplane transducer). The post-implant 3D ultrasound for each patient were inferred by CorneliUS for needle reconstruction as originally implemented—without retraining the model on local data. These AI reconstructions were compared to clinical needle reconstructions done by a trained operator at the time of implant. Differences in needle shaft reconstruction between the AI-reconstructed and clinical needle positions were quantified using mean Euclidian distance (MED). Additional metrics such as the percentage of implanted needles detected by CorneliUS were also compared to the tool’s original implementation.
Results
CorneliUS achieved a MED of (0.58±0.47) mm between the AI- and human-based reconstructions, with 90.7% of AI needle points within 1 mm of their human reconstruction counterparts. A two-tailed student t-test demonstrated this MED was statistically comparable to that of the original implementation of CorneliUS, (0.59±0.41) mm (p=0.74). The needle detection rate of our CorneliUS implementation (92.6%) was found to be similar to that of the original institution (2-sample z-test, p=0.93). Furthermore, the MED from our institution was smaller than the inter-operator MED between two human reconstructions reported at the original institution (0.64 ± 0.48 mm).
Conclusion
This study demonstrates the performance of CorneliUS in an independent clinical setting is very comparable to its original implementation, highlighting its potential to reduce the procedure time across multiple clinical sites. These results indicate the CorneliUS model is generalizable and does not required retraining when implemented in a similar clinical setting.