Deep convolutional neural networks for automatic segmentation of thoracic organs-at-risk in radiation oncology - use of non-domain transfer learning.

PURPOSE

Segmentation of organs-at-risk (OARs) is an essential component of the radiation oncology workflow. Commonly segmented thoracic OARs include the heart, esophagus, spinal cord, and lungs. This study evaluated a convolutional neural network (CNN) for automatic segmentation of these OARs.

METHODS

The dataset was created retrospectively from consecutive radiotherapy plans containing all five OARs of interest, including 22,411 CT slices from 168 patients. Patients were divided into training, validation, and test datasets according to a 66%/17%/17% split. We trained a modified U-Net, applying transfer learning from a VGG16 image classification model trained on ImageNet. The Dice coefficient and 95% Hausdorff distance on the test set for each organ was compared to a commercial atlas-based segmentation model using the Wilcoxon signed-rank test.

RESULTS

On the test dataset, the median Dice coefficients for the CNN model vs. the multi-atlas model were 71% vs. 67% for the spinal cord, 96% vs. 94% for the right lung, 96%vs. 94% for the left lung, 91% vs. 85% for the heart, and 63% vs. 37% for the esophagus. The median 95% Hausdorff distances were 9.5  mm vs. 25.3 mm, 5.1  mm vs. 8.1 mm, 4.0  mm vs. 8.0 mm, 9.8  mm vs. 15.8 mm, and 9.2 mm vs. 20.0 mm for the respective organs. The results all favored the CNN model (P < 0.05).

CONCLUSIONS

A 2D CNN can achieve superior results to commercial atlas-based software for OAR segmentation utilizing non-domain transfer learning, which has potential utility for quality assurance and expediting patient care.