Degree
Doctor of Philosophy (PhD)
Department
Physics & Astronomy
Document Type
Dissertation
Abstract
Despite the potential of deep learning (DL) to improve treatment planning workflow and plan quality, DL has not yet fully automated the treatment planning process. This study addresses that gap by applying DL to Gamma Knife (GK) treatment planning as a proof of concept. We focus on developing DL models that predict the number and location of non-overlapping shots for simple GK plans and explore using alternative contour volumes as inputs to improve coverage at the 50% isodose line. These efforts contribute to the long-term goal of generating clinically relevant GK plans through automation.
We trained and evaluated models using two architectures: standard U-Net and hierarchically dense (HD) U-Net. First, we developed models for 8mm non-overlapping shots, followed by models predicting both 4mm and 8mm shots. All models were trained and evaluated on randomly simulated GK targets. Predicted shot locations were extracted from model output using watershed segmentation, and dose distributions were calculated based on the resulting shot positions.
Model performance was assessed using the Dice Similarity Coefficient (DSC), 95% Hausdorff Distance (HD95), and deliverability metrics (coverage, selectivity, and gradient index). The 8mm-shot models achieved DSC > 0.97 and HD95 < 0.12 mm, while the mixed-shot models achieved DSC > 0.90 and HD95 < 2 mm. Deliverability metric errors were under 4% for the first study and under 10% for the second, confirming clinical relevance.
We also investigated input contour modifications to improve 50% isodose coverage. Strategy 1 used dilated contours (PTVs) as input, slightly improving mean coverage for some models (1–2%), but larger contour volumes with added margin > 0.5mm reduced coverage. Strategy 2 used binary contours based on 55% and 60% isodose levels to train new models. These models achieved higher mean coverage than the original 50% models when tested on GTVs, especially in mixed-shot predictions (up to 4% improvement). Thus, Strategy 2 proved more effective in improving coverage, particularly for plans involving both shot sizes.
Date
4-21-2025
Recommended Citation
Lay, Lam My, "GAMMA KNIFE PLAN PARAMETER PREDICTION WITH DEEP LEARNING METHODS" (2025). LSU Doctoral Dissertations. 6730.
https://repository.lsu.edu/gradschool_dissertations/6730
Committee Chair
David Solis
Included in
Artificial Intelligence and Robotics Commons, Health and Medical Physics Commons, Radiation Medicine Commons