High-dose-rate brachytherapy (HDRBT) is a radiation treatment for prostate cancer in which hollow needles are surgically inserted through the perineum into the prostate while the patient is under general anesthesia. A radioactive source -- Iridium-192 -- is then moved through each needle and dwells at calculated positions for precise amounts of time to deliver a highly concentrated dose directly to the tumor.
Before any dose can be delivered, each implanted needle must be precisely located in three dimensions using transrectal ultrasound (TRUS) images. This process, called needle digitization, converts the needle's spatial position in the ultrasound volume into a 3D mathematical description that the treatment planning software uses to calculate where and for how long the radiation source should pause inside each needle.
Errors in needle digitization translate directly into errors in the dose distribution: regions of the prostate may be underdosed while nearby organs -- the rectum, bladder, and urethra -- may receive more radiation than planned. Accurate digitization is therefore critical to both treatment effectiveness and patient safety.
Needle digitization is currently performed manually by a trained medical physicist or dosimetrist, a time-consuming and subjective process that takes approximately 10 minutes per patient and varies between operators depending on their experience. The patient remains under anesthesia throughout the entire treatment procedure, which can last close to 4 hours, making any reduction in procedure time directly beneficial to the patient.
The study used TRUS imaging data from 1102 prostate brachytherapy treatments performed at Orebro University Hospital between 2010 and 2019. Across these cases, a total of 24,422 individual needles had been manually digitized by medical physicists during actual treatments. These manual digitizations served as the clinical ground truth (CGT) for training and evaluation.
Each 3D TRUS volume was standardized to a fixed cubic size of 90 x 90 x 90 mm and then resampled to a uniform resolution of 128 x 128 x 128 voxels. Data were split into training (65%), validation (17.5%), and testing (17.5%) sets, yielding 17,215 needle examples for training and 7,207 for independent evaluation.
The network architecture was a 3D U-Net -- a fully convolutional encoder-decoder network widely used for volumetric medical image segmentation. The encoder path compressed the 3D input through four stages of 3x3x3 convolutions followed by max-pooling, reducing the spatial dimensions from 128x128x128 down to an 8x8x8 feature representation. The decoder path then progressively upsampled back to the original resolution using transposed convolutions, with skip connections transferring feature maps from each encoder level to the corresponding decoder level.
The network was optimized by maximizing the Dice similarity coefficient (DSC) between the predicted and manually labeled needle voxels. A dropout rate of 0.2 was applied during training to prevent overfitting. The Adam optimizer was used with a learning rate that decreased linearly from 1.2e-5 to 8e-6 over 50 training epochs, each taking 15 minutes on an NVIDIA GTX 2080 Ti GPU.
The 3D U-Net outputs a probability map in which each voxel receives a value between 0 and 1 indicating the likelihood that it is part of a needle. This continuous heatmap is then converted into a discrete set of clinically usable 3D needle descriptions through a geometric post-processing pipeline.
The probability map is first thresholded at 0.1 to identify candidate needle voxels. For each candidate needle and each imaging slice, the weighted centroid of predicted voxels in that slice is computed to yield a single (x, y) coordinate pair. Across all slices spanning the prostate, these coordinate pairs are then fitted using a least-squares linear regression, producing a 3D line equation that describes the needle's orientation and position throughout the prostate volume.
This linear representation matches the format used by clinical treatment planning systems, allowing the AI output to be directly imported into the dose planning workflow without additional manual steps. The comparison between AI and manual digitizations is then performed by measuring the root-mean-square deviation (RMSD) between the two 3D line representations, evaluated at 27 evenly spaced points spanning the 27 mm of the prostate along each needle.
Angular deviation (AD) -- the angle between the AI-predicted needle direction vector and the ground truth direction vector -- was also calculated as a complementary geometric accuracy measure. Both RMSD and AD are clinically meaningful because even a correctly localized needle tip combined with a wrong angle would shift dwell positions away from the planned locations inside the prostate.
Evaluated on 7,207 needles from 389 independent treatments never seen during training, the AI achieved a median RMSD of 0.55 mm (interquartile range 0.35 to 0.86 mm) relative to the clinical ground truth. The network identified every single needle in every treatment -- a 100% detection sensitivity -- meaning no needle was missed entirely.
To benchmark AI performance against human interobserver variability, a subset of 188 needles from 9 treatments was re-digitized by a second experienced medical physicist (designated G1). The difference between the original CGT and G1's re-digitization was 0.80 mm RMSD -- representing normal variability between two trained human operators. The AI's deviation from CGT on the same subset was 0.52 mm, statistically significantly smaller than either the AI-versus-G1 difference (0.75 mm) or the CGT-versus-G1 difference.
These results show that the AI's digitization style closely matches the original CGT operators rather than representing an independent perspective. The AI learned to replicate the specific conventions of the medical physicists who created the training data, which is clinically advantageous because it ensures consistency with the institutional treatment planning workflow.
The median angular deviation between AI and CGT was 0.88 degrees, compared to 1.00 degrees between CGT and G1. Both values indicate that the AI and the human operators agree on needle orientation to within approximately 1 degree, a clinically acceptable level of angular accuracy for treatment planning purposes.
Manual needle digitization by the medical physicist required a mean time of 10 minutes and 11 seconds per treatment. Since patients are under general anesthesia throughout the brachytherapy procedure -- which can take up to 4 hours -- even a 10-minute reduction in one step has meaningful clinical value by reducing anesthesia exposure time and operating room occupancy.
The AI completed the same digitization task in under 2 seconds, a reduction of more than 99% in time. If implemented in a clinical workflow, this would allow the medical physicist's role during treatment to shift from performing a repetitive digitization task to reviewing and verifying the AI output -- a supervisory role that is both faster and potentially more reliable than performing the task from scratch.
A key advantage of AI digitization compared to human digitization is consistency. Human operators differ subtly in how they interpret needle shadows and artifacts in ultrasound images, producing variation that is unrelated to true needle position. The AI, having learned from over 24,000 examples, applies the same pattern-matching process to every case without fatigue, distraction, or the learning curve that affects less experienced operators.
The study notes that accurate needle digitization is a prerequisite for accurate dose planning, and future work should quantify whether the small remaining positional differences between AI and manual digitization translate into clinically meaningful changes in the calculated dose distribution. If dose differences are negligible, the AI system could be implemented as a fully automated first step with human verification, directly reducing procedure time and operator burden.
The study's 0.55 mm RMSD compares favorably to prior published results. Wang et al., who also used a large training set of 823 patients and achieved 100% sensitivity, reported a higher RMSD of 0.74 mm. Other studies (Zhang et al. and Dise et al.) reported lower RMSD values of 0.29 and 0.40 mm but could not achieve 100% detection -- meaning some needles were missed entirely. Missing a needle in treatment planning is more clinically dangerous than a small positional error in a detected needle, making the present study's perfect sensitivity a notable strength.
A key methodological difference from prior work is the use of the Dice similarity coefficient as the training objective. Earlier needle detection studies typically used weighted combinations of cross-entropy and total variation losses trained on image patches rather than whole volumes. The whole-volume approach used here preserves spatial context across the entire prostate, which may contribute to the improved detection rate.
The study has several acknowledged limitations. The network was not trained to find needle tips, which in clinical practice are located using a separate x-ray measurement rather than TRUS. Incorporating tip detection would require a different training approach or supplementary data. The RMSD evaluation was only performed within the prostate volume (27 mm along each needle), not along the full needle length, meaning accuracy outside the treatment region was not assessed.
The algorithm was trained exclusively on images from two specific ultrasound systems, and its performance on different equipment is unknown. Generalization to other institutions and scanner types will require either retraining or transfer learning approaches. The accuracy of the CGT itself is also limited by the resolution and contrast of TRUS, which means the ground truth contains measurement uncertainty that cannot be separated from the AI's actual error.