Related Experiment Video
Updated: Jun 25, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.7K
Enhancing Precision in Cardiac Segmentation for Magnetic Resonance-Guided Radiation Therapy Through Deep Learning
Nicholas Summerfield1, Eric Morris2, Soumyanil Banerjee3
1Department of Medical Physics, University of Wisconsin-Madison, Madison, Wisconsin; Department of Human Oncology, University of Wisconsin-Madison, Madison, Wisconsin.
Summary
This study introduces nnU-Net.wSD, a deep learning model that accurately segments cardiac substructures for improved radiation therapy. This advancement enhances cardiac sparing during Magnetic Resonance-guided Radiation Therapy (MRgRT).
Area of Science:
- Medical Physics
- Radiotherapy
- Artificial Intelligence
Background:
- Cardiac substructure dose metrics correlate strongly with late cardiac morbidities.
- Magnetic Resonance-guided Radiation Therapy (MRgRT) allows for daily visualization of cardiac substructures.
- Accurate segmentation of cardiac substructures is crucial for effective cardiac sparing in radiation therapy.
Purpose of the Study:
- To extend the "No New" U-Net deep learning framework with self-distillation (nnU-Net.wSD) for improved cardiac substructure segmentation in MRgRT.
- To evaluate the performance of nnU-Net.wSD compared to standard 3D U-Net for cardiac substructure segmentation.
- To assess the impact of data augmentation strategies on the segmentation accuracy of nnU-Net.wSD.
Main Methods:
- Retrospective evaluation of 18 patients undergoing thoracic or abdominal radiation therapy on a 0.35 T MR-guided linear accelerator.
- Delineation of 12 cardiac substructures by radiation oncologists for training, validation, and testing of nnU-Net.wSD.
- Comparison of nnU-Net.wSD with 3D U-Net using geometric metrics (Dice, MDA, Hausdorff) and analysis of dose-volume histograms.
- Evaluation of generalizability using an independent dataset from a second institute.
Main Results:
- nnU-Net.wSD achieved a mean Dice similarity coefficient of 0.65 ± 0.25 across 12 cardiac substructures, outperforming 3D U-Net (0.583 ± 0.28, P < .01).
- Augmentation with fractionated data significantly improved performance over single MR simulation time points (P < .01).
- Predicted contours resulted in dose-volume histograms closely matching clinical plans, with minimal deviations in mean (0.32 ± 0.5 Gy) and maximum doses (1.42 ± 2.6 Gy).
Conclusions:
- nnU-Net.wSD demonstrates robust performance for cardiac substructure segmentation in MRgRT.
- The developed framework is a significant step towards rapid and reliable cardiac substructure segmentation.
- This technology has the potential to substantially improve cardiac sparing during MRgRT, reducing late cardiac morbidities.
