Related Experiment Video
Updated: Aug 2, 2026
12:41
Multiparametric Tumor Organoid Drug Screening Using Widefield Live-Cell Imaging for Bulk and Single-Organoid Analysis
Published on: December 23, 2022
4.7K
Point-cloud segmentation with in-silico data augmentation for prostate cancer treatment
Jianxin Zhou1, Massimiliano Salvatori2, Kadishe Fejza3
1Department of Nuclear, Plasma, and Radiological Engineering, University of Illinois at Urbana-Champaign, Urbana, Illinois, USA.
Medical Physics
|April 4, 2025
Summary
This study introduces a fast point-cloud segmentation model for adaptive radiation therapy (ART). The novel approach significantly reduces organ segmentation time in CT scans, improving treatment workflow efficiency and accuracy for prostate cancer patients.
Area of Science:
- Medical Physics and Imaging
- Radiotherapy and Oncology
- Artificial Intelligence in Medicine
Background:
- External x-ray radiation therapy dose distributions can be inaccurate due to patient positioning, anatomical changes, and delivery system limitations.
- Adaptive radiation therapy (ART) offers improved accuracy by re-optimizing treatment plans, but requires rapid image segmentation for feasibility.
- Faster segmentation is crucial for routine ART implementation, enabling personalized treatment, dose escalation, and reduced toxicity.
Purpose of the Study:
- To develop a rapid, point-cloud-based segmentation model for pelvic CT scans in prostate cancer (PCa) treatment.
- To incorporate novel in-silico data augmentation for enhanced model training and robustness.
- To demonstrate the model's potential for real-time implementation within ART workflows.
Main Methods:
- A novel deep learning point-cloud-based network was developed, featuring a combined region-based and boundary loss function for accurate organ segmentation.
- In-silico data augmentation using the XCAT phantom generated synthetic 3D CT images to expand the training dataset.
- The model was trained and validated on pelvic CT data, with performance assessed using the Dice similarity coefficient and segmentation time compared to existing methods.
Main Results:
- The point-cloud model achieved high segmentation accuracy (Dice coefficients: Bladder 0.92, Prostate 0.89, Rectum 0.84).
- Segmentation time was significantly reduced: 1.8x faster than 2D FCN and 11x faster than 3D U-Net.
- The model demonstrated robustness across diverse images, outperforming voxel-based methods in prostate segmentation accuracy.
Conclusions:
- The proposed point-cloud segmentation method is faster and achieves comparable or superior results to voxel-based algorithms for CT data.
- The model's speed and accuracy make it suitable for integration into ART workflows, potentially reducing clinician workload.
- This approach supports efficient and personalized radiation therapy, improving cancer treatment outcomes and patient safety.
