Training and Validation of Deep Learning-Based Auto-Segmentation Models for Lung Stereotactic Ablative Radiotherapy...

Area of Science:

• Radiotherapy
• Medical Imaging
• Artificial Intelligence

Background:

• Deep learning (DL) auto-segmented contour (DC) models require high-quality data.
• Prospectively produced contours are resource-intensive and time-consuming.
• Retrospective data offers a potential alternative for DC model development.

Purpose of the Study:

• To assess the feasibility of using retrospective peer-reviewed radiotherapy planning contours.
• To train and evaluate deep learning contouring models for lung stereotactic ablative radiotherapy (SABR).

Main Methods:

• Trained DL contouring models using retrospective contours from 160 public CT scans and 50 peer-reviewed SABR 4D-CT scans.
• Generated contours for 50 additional CT scans from two centers.
• Compared auto-segmented contours (DCs) with clinical contours (CCs) using Dice Similarity Coefficient (DSC) and 95% Hausdorff distance (HD).

Main Results:

• Mean DSC and 95% HD varied by organ at risk (OAR), with high accuracy for aorta (0.93 DSC, 2.85mm HD) and spinal cord (0.90 DSC, 1.62mm HD).
• Gross tumor volume (GTV) showed lower accuracy (0.71 DSC, 5.23mm HD).
• Contour accuracy was consistent across different centers.

Conclusions:

• Deep learning contouring models trained on retrospective data approximate clinical contours for most OARs.
• Structures with higher variability may require more training data or novel approaches.
• Developing DL models from existing radiotherapy planning contours is feasible and warrants further clinical workflow integration.