Application of a multi-task learning-based light-weight convolution neural network for the automatic segmentation of organs at risk in thorax
10.3760/cma.j.cn113030-20200409-00167
- VernacularTitle:基于多任务学习的轻量级卷积神经网络在肺部危及器官自动分割中的应用研究
- Author:
Jie ZHANG
1
;
Yiwei YANG
;
Kainan SHAO
;
Xue BAI
;
Min FANG
;
Guoping SHAN
;
Ming CHEN
Author Information
1. 中国科学院大学附属肿瘤医院（浙江省肿瘤医院）放射物理科，杭州 310022
- Keywords:
Multi-task learning;
Light-weight convolution neural network;
Automatic segmentation;
Organs at risk
- From:
Chinese Journal of Radiation Oncology
2021;30(9):917-923
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To evaluate the application of a multi-task learning-based light-weight convolution neural network (MTLW-CNN) for the automatic segmentation of organs at risk (OARs) in thorax.Methods:MTLW-CNN consisted of several layers for sharing features and 3 branches for segmenting 3 OARs. 497 cases with thoracic tumors were collected. Among them, the computed tomography (CT) images encompassing lung, heart and spinal cord were included in this study. The corresponding contours delineated by experienced radiation oncologists were ground truth. All cases were randomly categorized into the training and validation set ( n=300) and test set ( n=197). By applying MTLW-CNN on the test set, the Dice similarity coefficients (DSCs) of 3 OARs, training and testing time and space complexity (S) were calculated and compared with those of Unet and DeepLabv3+ . To evaluate the effect of multi-task learning on the generalization performance of the model, 3 single-task light-weight CNNs (STLW-CNNs) were built. Their structures were totally the same as the corresponding branches in MTLW-CNN. After using the same data and algorithm to train STLW-CNN, the DSCs were statistically compared with MTLW-CNN on the testing set. Results:For MTLW-CNN, the averages (μ) of lung, heart and spinal cord DSCs were 0.954, 0.921 and 0.904, respectively. The differences of μ between MTLW-CNN and other two models (Unet and DeepLabv3+ ) were less than 0.020. The training and testing time of MTLW-CNN were 1/3 to 1/30 of that of Unet and DeepLabv3+ . S of MTLW-CNN was 1/42 of that of Unet and 1/1 220 of that of DeepLabv3+ . The differences of μ and standard deviation (σ) of lung and heart between MTLW-CNN and STLW-CNN were approximately 0.005 and 0.002. The difference of μ of spinal cord was 0.001, but σof STLW-CNN was 0.014 higher than that of MTLW-CNN.Conclusions:MTLW-CNN spends less time and space on high-precision automatic segmentation of thoracic OARs. It can improve the application efficiency and generalization performance of the models.
