Unsupervised Deep Learning Features for Lung Cancer Overall Survival Analysis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Lung cancer survival analysis is crucial for treatment planning.
• Current methods rely on hand-crafted features, which require expertise and may lack specificity.
• Deep learning models show promise but need large datasets, challenging for survival analysis due to long follow-up times.

Purpose of the Study:

• To develop an unsupervised deep learning method for lung cancer survival analysis that leverages unlabeled data.
• To compare the performance of deep learning features against traditional hand-crafted features.

Main Methods:

• Proposed a residual convolutional autoencoder for unsupervised feature learning from unlabeled computed tomography (CT) images.
• Trained the autoencoder on 274 lung cancer patients without survival data.
• Extracted deep learning features and applied a Cox proportional hazards model to 129 patients with survival data.

Main Results:

• The unsupervised deep learning features achieved a higher C-Index (0.70) compared to hand-crafted features (0.62).
• Kaplan-Meier analysis demonstrated the model's ability to stratify patients into distinct high and low-risk groups.
• The survival times between the identified risk groups showed a statistically significant difference (p < 0.01).

Conclusions:

• Unsupervised deep learning offers a viable solution for lung cancer survival analysis, especially when labeled data is scarce.
• The proposed method effectively extracts relevant features from CT images for survival prediction.
• This approach enhances the ability to predict patient risk stratification and survival outcomes.