Room: AAPM ePoster Library
We developed a semi-supervised learning method using soft-label for lung segmentation on CT.
The proposed segmentation method was applied on Lung CT Segmentation Challenge 2017. This data set comprises of average 4DCT or free-breathing (FB) CT images from 60 patients, depending on clinical practice. 2D images were extracted from the datasets and were divided into three groups: 70 training images, 20 validation images, 10 test images. Within the training images, 35 images were used as labeled data and 35 images were used as unlabeled data.
A renovated mean-teacher semi-supervised framework was developed which encouraged the segmentation predictions to be consistent during different noises for the same input. The framework included two identical architecture networks which were a student model and a teacher model. The student model was regularized by a consistency soft loss with the teacher model and a supervised segmentation loss with the labels. After the weights of the student model were updated with gradient descent, the weights of the teacher model were updated as an exponential moving average (EMA) of the student weights.
The quantitative results of our segmentation method achieved mean dice score of (0.96, 0.98), mean accuracy of (0.9983, 0.9984), and mean relative error of (0.037, 0.071) with 95% CI on the 10 test datasets.
The qualitative and quantitative comparisons show that our proposed method can achieve higher segmentation accuracy with less variance on testing datasets. It will be useful in image analysis applications for lung lesions diagnosis and radiotherapy assessment in thoracic CT.
Segmentation, Image Analysis, Computer Vision