2022 World Conference on Lung Cancer (Posters)-P1.10-03. A Deep Learning Auto-Segmentation Tool for Cardiac Substructures in 4D Radiotherapy Planning for Locally Advanced Lung Cancer

P1.10-03. A Deep Learning Auto-Segmentation Tool for Cardiac Substructures in 4D Radiotherapy Planning for Locally Advanced Lung Cancer

Back to course

Pdf Summary

This study aimed to validate a deep learning-based auto-segmentation tool for cardiac substructures in 4D radiotherapy planning scans. The tool was previously trained on 240 lung cancer treatment plans to generate whole heart, pericardium, and 10 cardiac substructures. The researchers hypothesized that the tool's performance on 4D average intensity projection (4D-AVE) scans would be comparable to 3D-CT scans using various metrics.<br /><br />The study included a dataset of 20 patients who underwent radical radiotherapy for lung cancer. Manual and automated cardiac segmentation were compared using volume difference, centroid shift, dice similarity coefficient, and Hausdorff distance. The mean and maximum doses to the structures were also compared.<br /><br />The results showed that automated volumes were generally smaller for the chambers and larger for the great vessels compared to manual delineations. The volume differences were statistically significant for some structures. The dice similarity coefficient values ranged from 0.76 to 0.94, with the lowest performance observed for the inferior vena cava and the best for the pericardium. The centroid shift values were approximately 3-4 mm for most structures. The Hausdorff distance values were small, indicating good agreement between the automated and manual structures. The mean and maximum doses to the structures were similar for both automated and manual segmentations.<br /><br />Overall, the automated contours were deemed clinically acceptable by two senior clinicians, and minor modifications led to insignificant changes in geometry. The study concluded that the deep learning-based tool had high performance on 4D-CT planning scans and could be used for cardiac substructure segmentation in clinical practice.<br /><br />One limitation of the study was the small cohort size. However, the results were considered clinically insignificant compared to the original 3D-CT publication. The study highlighted the potential of deep learning-based tools for accurate and efficient cardiac segmentation in radiotherapy planning.

Asset Subtitle

Gerard Walls, United Kingdom

Meta Tag

Speaker Gerard Walls, United Kingdom

Topic Locally Advanced Non-small Cell Lung Cancer - Chemoradiotherapy and Radiotherapy

Keywords

deep learning

auto-segmentation

cardiac substructures

4D radiotherapy planning scans

lung cancer treatment plans

whole heart

pericardium

cardiac substructure segmentation

3D-CT scans

radiotherapy planning