Low-dose rate (LDR) brachytherapy is an established treatment for early-stage prostate cancer in which tiny radioactive seeds are permanently implanted directly inside the prostate. The seeds emit radiation continuously at low energy, delivering a precisely localized dose that destroys cancer cells while sparing surrounding tissue -- achieving excellent local tumor control and preserving quality of life.
Traditionally, LDR brachytherapy for the prostate uses a fixed grid template mounted between the patient's legs, through which needles are inserted in a parallel, coplanar pattern under transrectal ultrasound guidance. While effective, this grid-based approach restricts needle trajectories to predefined paths and limits the physician's ability to reach awkwardly shaped or positioned tumor volumes while avoiding sensitive structures like the urethra, rectum, and pubic arch.
Robotic assistance offers a new degree of freedom: a robotic arm can automatically position an injection template at any point on the patient's skin surface with sub-millimeter precision, enabling needles to approach the prostate from multiple directions simultaneously. This multi-directional capability has the potential to improve dose conformality -- more accurately shaping the radiation dose to match the tumor contour -- and reduce the risk of under-dosing difficult regions.
However, this expanded freedom creates a new computational challenge: with needles arriving from many possible angles, the number of possible needle placement combinations becomes enormous. A sophisticated treatment planning system (TPS) is needed to efficiently identify which needle paths are safe, and then optimize the placement of radioactive seeds along those paths to achieve the desired dose distribution.
The system was tested and validated using a complex abdominal phantom -- a physical model of the human abdomen containing a simulated liver metastasis (40.3 ml volume) surrounded by anatomical structures including the liver, aorta, vena cava, and ribs. This liver metastasis scenario was deliberately chosen as a challenging test case to demonstrate the system's ability to handle tumor sites beyond the prostate.
The workflow begins with a cone-beam CT (CBCT) scan of the phantom using a robotic angiography system. Oncologists then contour the target volume (the metastasis) and all risk structures on the CT images. These contoured structures are imported into the custom-built treatment planning system, implemented in MATLAB, which then executes all planning and optimization steps.
The prescribed dose for the liver metastasis test case was 100 Gy delivered to the covering isodose surface. Planning targets included achieving V100 greater than 99% (meaning at least 99% of the target volume receives the full prescription dose), while minimizing the volume receiving 200% of the dose (V200), and achieving a D90 greater than 120% of the prescription dose. D90 is the minimum dose received by 90% of the target volume and is a key predictor of treatment success in brachytherapy.
The path planning algorithm begins by identifying all points on the patient's skin surface as candidate needle injection sites. It then systematically filters these candidates using a multi-step screening process that eliminates any trajectory passing through risk structures such as the liver, ribs, aorta, or vena cava. This filtering process eliminated 72.5% of all initially considered injection points, leaving a manageable set of safe trajectories.
For each valid injection point, the algorithm calculates all reachable positions within the target volume (tip points), forming a candidate needle domain -- the complete set of all geometrically possible needle trajectories. In the phantom test, this domain contained 1,971 tip points and 827 injection points. This comprehensive domain is then passed to the dose optimization stage.
A key technical challenge is handling irregularly shaped structures like the liver, whose complex boundary makes simple geometric filtering insufficient. The algorithm addresses this by computing a standard deviation vector for each risk structure that captures its three-dimensional extent in each direction, enabling more accurate containment checks that account for structural asymmetry.
Once the candidate needle domain is established, a modular four-step inverse optimization algorithm generates the final treatment plan. The first step, the greedy optimizer, builds an initial plan by sequentially adding needles and seeds until at least 99% of the target volume receives the full prescription dose. Needles are added one by one, each time selecting the trajectory that most improves coverage.
The second step, the remove-seed algorithm, identifies and removes seeds located in regions that already receive far more than the prescribed dose (hotspots). Removing these excess seeds reduces the V200 by approximately 15% while decreasing V100 by only about 1% -- a favorable tradeoff that improves dose conformality without compromising coverage.
The third step, the depth optimizer, fine-tunes the insertion depth of each needle to further optimize the distribution of dose within the target. The fourth step, the coverage optimizer, identifies any remaining under-dosed regions and adds supplementary needles to ensure complete coverage. Together these four steps balance the competing goals of complete tumor coverage, minimizing hotspots, and using the fewest possible needles and seeds.
Across 10 independent planning runs on the same phantom, the system consistently achieved excellent dose coverage: V100 = 99.1 +/- 0.3%, meaning virtually the entire target received the full prescription dose. V150 was 76.4% and V200 was 44.5%, indicating acceptable dose distribution. The D90 of 125.9 Gy comfortably exceeded the 120% threshold associated with good clinical outcomes.
Plans used an average of 10.7 needles containing 34 seeds, comparable to standard prostate brachytherapy. Total TPS running time had a median of 4.4 minutes, which is clinically acceptable for an intraoperative or pre-operative planning scenario. In two outlier cases, optimization took longer (up to 29 minutes) due to the coverage optimizer requiring repeated iterations to satisfy dose constraints in geometrically challenging configurations.
Dose accuracy was validated by comparing the custom TPS dose calculation to a commercially approved and CE-marked treatment planning system (Oncentra Prostate). A global gamma analysis -- a standard radiation oncology metric for dose agreement -- showed a pass rate of 98.5% at 1% dose and 1 mm distance thresholds, confirming that the custom dose calculation is accurate enough for clinical use.
This work demonstrates a complete, end-to-end robotic brachytherapy planning pipeline that is fast enough for clinical use and accurate enough for safe patient treatment. By enabling needles to approach the tumor from multiple directions, the system overcomes the limitations of traditional template-based brachytherapy and could improve outcomes for tumors near challenging anatomical structures.
Beyond the prostate, the system's generalizability is a major strength. The same framework was tested on a liver metastasis scenario, demonstrating potential applicability to a wide range of tumor sites where LDR brachytherapy has historically been limited by the inability to navigate safely around complex risk structures using conventional parallel needle approaches.
Future development priorities include further validation in patient-specific anatomies with more complex geometries, prospective clinical studies comparing multi-directional robotic brachytherapy to standard template-based approaches, and integration with automated anatomical contouring tools to streamline the planning workflow from imaging to treatment delivery.