Improving accuracy and robustness of deep convolutional neural network based thoracic OAR segmentation.

Researchers

Journal

Modalities

Models

Abstract

Deep convolutional neural network (DCNN) has shown great success in medical image segmentation. However, most studies involve training and testing on the same dataset. Little work has been done to investigate the generalization errors of DCNN on a different dataset. This work investigated the generalization errors observed when apply a well-trained DCNN model to data from our local institution. It was found that even a subtle shift of organs inside patient body, caused by abdominal compression technique used in our institution, can confuse a well-trained DCNN. Incorporating cases from local institution into training process improves the accuracy and robustness of the DCNN model. The number of local cases were varied to see the impact on the performance improvement. In addition, we also picked cases that the original DCNN performed poorly for training to see if DCNN can learn more effectively from it’s mistakes. Our finding is that the DCNN can learn this change in anatomy with only a few cases without extra effort in data collection.
© 2020 Institute of Physics and Engineering in Medicine.