Abstract
In radiotherapy for cancer patients, an indispensable process is to delineate organs-at-risk (OARs) and tumors. However, it is the most time-consuming step as manual delineation is always required from radiation oncologists. Herein, we propose a lightweight deep learning framework for radiotherapy treatment planning (RTP), named RTP-Net, to promote an automatic, rapid, and precise initialization of whole-body OARs and tumors. Briefly, the framework implements a cascade coarse-to-fine segmentation, with adaptive module for both small and large organs, and attention mechanisms for organs and boundaries. Our experiments show three merits: 1) Extensively evaluates on 67 delineation tasks on a large-scale dataset of 28,581 cases; 2) Demonstrates comparable or superior accuracy with an average Dice of 0.95; 3) Achieves near real-time delineation in most tasks with <2 s. This framework could be utilized to accelerate the contouring process in the All-in-One radiotherapy scheme, and thus greatly shorten the turnaround time of patients.
Introduction
Precision is a double-edged sword in modern radiotherapy. To effectively target a tumor while sparing healthy organs-at-risk (OARs), clinicians must meticulously trace boundaries on 3D CT scans. This process, known as volume delineation or contouring, is the primary bottleneck in cancer treatment, often taking hours or even days per patient. Furthermore, manual contouring is prone to inter-observer variability, where different oncologists may delineate the same structure differently.
While deep learning models like U-Net have shown promise in automating this task, they are often computationally intensive, requiring significant GPU memory and minutes of processing time. This latency hinders the goal of "All-in-One" radiotherapy—a paradigm where a patient is diagnosed, planned, and treated in a single clinical visit.
Methods: The RTP-Net Framework
RTP-Net is designed for extreme efficiency without sacrificing segmentation precision. Its architecture incorporates several key innovations:
- Cascade Coarse-to-Fine Strategy: Instead of processing high-resolution 3D volumes directly, the system first employs a low-resolution model to localize a "Region of Interest" (ROI). A subsequent fine-resolution model then segments the specific boundaries. This approach drastically reduces the processing of irrelevant data and minimizes GPU memory requirements.
- Adaptive Input Module: Anatomical structures vary significantly in scale. RTP-Net utilizes a flexible VB-Net (a modified V-Net) architecture that dynamically adjusts down-sampling layers based on the target organ's size.
- Dual Attention Mechanisms: To improve boundary precision, the model incorporates "OAR-aware" and "boundary-aware" attention maps. These act as internal guides, focusing the network on critical edge voxels.
- Multi-dimensional Adaptive Loss: The training process utilizes a custom loss function that emphasizes 2D slice boundaries, where segmentation errors are most prevalent.
Experimental Results
The study's scale is unprecedented, utilizing a dataset of 28,581 cases covering 67 different delineation tasks. Data was sourced from major public benchmarks (TCIA, SegTHOR, CHAOS, MSD, LUNA) and a local dataset of 362 rectal cancer images from Fudan University Shanghai Cancer Center.
Accuracy and Speed
RTP-Net achieved an average Dice Similarity Coefficient (DSC) of 0.95, performing comparably to or better than "heavyweight" state-of-the-art models like nnU-Net and Swin UNETR. The standout metric, however, is speed: RTP-Net completes most tasks in less than 2 seconds, whereas traditional high-performance models like nnU-Net can take over 200 seconds.
Clinical Impact
In practical applications, such as rectal cancer treatment planning, the model successfully delineated the Clinical Target Volume (CTV) and Planning Target Volume (PTV) with high precision, even at challenging upper and lower boundaries.
Discussion and Conclusion
RTP-Net demonstrates that optimizing for efficiency can enable entirely new clinical paradigms. By reducing contouring time from hours to seconds, it facilitates "in-situ" treatment planning. While certain structures like mediastinal lymph nodes remain challenging due to blurry boundaries (Dice 0.61), the overall framework significantly advances the "All-in-One" radiotherapy vision. Future work will focus on integrating RTP-Net into RT-linac platforms to further shorten the patient turnaround time from days to minutes.
Summary updated