TL;DR
This paper introduces a novel automated radiotherapy planning system for cervical cancer that leverages deep learning to predict dose distributions and then optimizes these predictions using "dose rings" within standard treatment planning systems (TPSs). This approach successfully generates clinically deliverable plans, significantly reducing planning time while maintaining or improving plan quality compared to manual methods.
What Problem Does This Paper Address?
Radiotherapy planning, particularly for complex techniques like Volumetric Modulated Arc Therapy (VMAT) and Intensity-Modulated Radiation Therapy (IMRT), is a time-consuming and iterative process. While deep learning (DL) has shown promise in predicting dose distributions, a major hurdle has been translating these predicted distributions into clinically deliverable plans within existing treatment planning systems (TPSs). Existing DL approaches often produce voxel-based plans that are not readily importable or optimizable in commercial TPSs, or they are limited to specific TPSs, hindering universal applicability. This paper aims to bridge this gap by developing a universally applicable and deliverable automated planning method that integrates DL dose prediction with a novel dose ring optimization strategy.
Methods
The core of this work lies in a two-stage automated planning process:
- Deep Learning Dose Prediction: A 3D Fusion Residual Unet (F-ResUnet) model was employed to predict the 3D dose distributions. This network architecture is designed to fuse information from different levels of abstraction, enabling it to capture complex spatial relationships within the patient's anatomy and the radiation dose. The model was trained on a dataset of VMAT and IMRT plans from gynecological cancer patients.
- Dose Rings Optimization: The predicted dose distributions were then discretized into "dose rings." These rings represent specific dose ranges (e.g., 1-2 Gy, 2-3 Gy, etc.). By defining constraints on these dose rings within standard commercial TPSs (Monaco and Pinnacle), the system automatically optimizes the plan. This approach leverages the spatial information from the DL prediction while utilizing the robust optimization engines of existing TPSs to ensure deliverability.
Datasets
The study utilized retrospective data from two hospitals:
- VMAT Patients: 230 patients undergoing VMAT for gynecological cancer.
- IMRT Patients: 210 patients undergoing IMRT for gynecological cancer.
These datasets were divided into training, validation, and testing sets. The study focused on cervical cancer patients, with some endometrial cancer patients included. No new benchmark datasets were introduced; rather, existing clinical plans served as the ground truth.
Key Insights and Ideas
- Bridging the DL-to-TPS Gap: The "dose rings" concept is a key innovation. By converting DL-predicted dose distributions into a generalized format (dose rings) that can be directly used as constraints in commercial TPSs, the method overcomes a significant barrier to clinical implementation.
- Universality and Deliverability: The dose ring approach aims for universality, meaning the same constraint functions can be applied across different TPSs, and the optimization within the TPS ensures the generated plans are clinically deliverable.
- Efficiency Gains: The automated process significantly reduces planning time, freeing up dosimetrists for more complex tasks or patient care.
- Balancing Target Coverage and OAR Sparing: The method demonstrates the ability to achieve comparable target coverage to manual plans while often improving organ-at-risk (OAR) sparing.
Key Results
- Clinical Acceptability: Predicted dose distributions were clinically acceptable, with target coverage exceeding 95%.
- Target Coverage: Automatic plans achieved an average target coverage of over 96.5%, comparable to clinical plans.
- OAR Sparing: Automatic VMAT plans significantly reduced the V30Gy for the left and right femoral heads and small intestine. Automatic IMRT plans substantially reduced V45Gy for the bladder and rectum (7% and 9% reductions, respectively).
- Deliverability: Patient-specific quality assurance (PSQA) showed high gamma passing rates (95.0% to 99.1%) under standard criteria, confirming the deliverability of the automated plans.
- Time Savings: The automatic planning process took significantly less time than manual planning. For VMAT, the automatic time was 11.38 minutes compared to 25.22 minutes manually. For IMRT, it was 5.07 minutes versus 17.35 minutes.
Limitations and Future Work
- Dataset Size: The study acknowledges that the DL models were trained on a relatively small number of cases. Further validation with larger, multi-institutional datasets is recommended.
- Limited Patient Population: The study focused on gynecological cancer patients. Extending the method to other cancer sites, such as head-and-neck cancers, is suggested.
- Hotspot Analysis: The current method does not explicitly consider the spatial distribution of dose hotspots within the automatic plan. Future work could investigate this aspect.
- External Validation: Cases from other hospitals are needed to further validate the models externally.
Summary updated