Enhancing HDR Prostate Brachytherapy Deep Learning Needle Reconstruction with Live Ultrasound-Assisted Instance Detection
Abstract
Purpose
We previously developed an automatic catheter reconstruction model (CorneliUS) that achieves approximately 93% needle detection efficiency in post‑implant 3D ultrasound images for prostate HDR brachytherapy. The remaining undetected (AI‑missed) needles are typically associated with image artefacts, making them difficult to identify even by a human operator. This work aims to develop a machine‑learning tool for real‑time catheter detection on live ultrasound images during implantation, providing positional estimates for AI‑missed needles.
Methods
A real‑time needle‑detection neural network based on the Ultralytics YOLO11 framework, CorneliOLO, was trained using manually labeled ultrasound video frames from 13 HDR prostate brachytherapy procedures. In the proposed workflow, needle instances are automatically detected in the live ultrasound feed as the implant progresses. If a needle is missing from the CorneliUS reconstruction on post‑implant 3D ultrasound, its last visible live‑feed position—recorded by CorneliOLO—is used to estimate its final location. This estimation leverages the relative spatial relationships of adjacent needles detected by both models as anchors within an affine registration framework. The method was retrospectively evaluated on 15 implants with 1–3 AI‑missed needles. Accuracy was quantified using the mean Euclidean distance (MED) between estimated positions and ground truth manual clinical segmentation.
Results
The proposed method achieved a MED of (1.34±0.65) mm for AI-missed needles, with 91% reconstructed within 2mm. Although this exceeds the (0.59±0.41) mm MED reported for AI‑detected needles using CorneliUS, integrating CorneliOLO effectively increases overall needle detection from ~93% to near‑complete, as CorneliOLO consistently provides positional estimates for needles missed on post‑implant imaging.
Conclusion
Integrating live ultrasound data during implantation enables reliable estimation of obscured needle positions on post‑implant images, yielding a more complete reconstruction pipeline than CorneliUS alone. This approach also lays foundational work for real‑time adaptive prostate brachytherapy dosimetry using live ultrasound imaging.