Unsupervised Deep Learning Features for Lung Cancer Overall Survival Analysis.

Abstract

Lung cancer overall survival analysis using computed tomography (CT) images plays an important role in treatment planning. Most current analysis methods involve hand-crafted image features for survival time prediction. However, hand-crafted features require domain knowledge and may lack specificity to lung cancer. Advanced self-learning models such as deep learning have showed superior performance in many medical image tasks, but they require large amount of data which is difficult to collect for survival analysis because of the long follow-up time. Although data with survival time is difficult to acquire, it is relatively easy to collect lung cancer patients without survival time. In this paper, we proposed an unsupervised deep learning method to take advantage of the unlabeled data for survival analysis, and demonstrated better performance than using hand-crafted features. We proposed a residual convolutional auto encoder and trained the model using images from 274 patients without survival time. Afterwards, we extracted deep learning features through the encoder model, and constructed a Cox proportional hazards model on 129 patients with survival time. The experiment results showed that our unsupervised deep learning feature gained better performance (C-Index = 0.70) than using hand-crafted features (C-Index = 0.62). Furthermore, we divided the patients into two groups according to their Cox hazard value. Kaplan-Meier analysis indicated that our model can divide patients into high and low risk groups and the survival time of these two groups had significant difference (p < 0.01).