Automatic Multi-Organ Segmentation Using a Deep Neural Network for Assessing Dose to Organs at Risk During Breast Radiotherapy
M Saha1*, J Jung2, S Lee3, C Lee4, C Lee1, M Mille1, (1) National Cancer Institute, Bethesda, MD (2) East Carolina Univ, Greenville, NC, (3) University of Maryland School of Medicine, Baltimore, MD, (4) University of Michigan, Ann Arbor, MI
Presentations
SU-H400-IePD-F6-4 (Sunday, 7/10/2022) 4:00 PM - 4:30 PM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 6
Purpose: To demonstrate a deep neural network method for the automatic segmentation of heart, aorta, trachea, and esophagus on breast radiotherapy planning computed tomography (CT) images.
Methods: A three-dimensional (3D) deep neural network based on nnUnet was created using a full resolution setup to automatically segment the heart, aorta, trachea, and esophagus on CT. The neural network was trained on a competition dataset which included 40 low-contrast thoracic CT images with manual annotations for the four organs. Segmentation performance was evaluated using five-fold Monte Carlo cross-validation and Dice Similarity Coefficient (DSC). The trained network was then applied to an independent set of 60 left breast radiotherapy planning CT images and the accuracy of the organ segmentation was compared to manual segmentations in terms of DSC and calculated dose.
Results: Our method demonstrated a remarkable ability to segment the four organs on the CT images achieving DSC values of 0.94, 0.93, 0.89, and 0.82 for the heart, aorta, trachea, and esophagus, respectively. Similar performance was observed when the network was applied to the breast radiotherapy plans. The mean difference in mean heart dose for the manual and automatic contours was 0.35 Gy ± 1.67 Gy (for prone setup) and -1.18 Gy ± 1.80 Gy (for supine setup). The overall (prone + supine) average of dose value is -0.4 Gy ± 1.9 Gy.
Conclusion: Our deep neural network segmentation method demonstrated excellent performance and can be extended to other organs to support the development of improved prescriptive criteria for minimizing late health effects after breast radiotherapy.
Keywords
Not Applicable / None Entered.
Taxonomy
IM- CT: Machine learning, computer vision
Contact Email