A deep attention LSTM embedded aggregation network for multiple histopathological images.

Area of Science:

• Medical imaging analysis
• Computational pathology
• Deep learning for survival prediction

Background:

• Deep learning advances medical imaging survival analysis.
• Current methods struggle with multiple lesion images per patient, complicating predictions.
• A patient-level prediction model is needed for improved interpretability.

Purpose of the Study:

• To develop a deep learning survival model for accurate patient-level predictions from multiple histopathology images.
• To address the challenge of interpreting multiple survival predictions from individual lesions.
• To create a comprehensive survival model that aggregates lesion-level information effectively.

Main Methods:

• Proposed a deep attention long short-term memory embedded aggregation network (DALAN).
• DALAN utilizes a weight-shared CNN for feature extraction and attention/LSTM layers for lesion image aggregation.
• The model learns imaging features and aggregates lesion information to the patient level.

Main Results:

• DALAN demonstrated superior prediction accuracy on simulated and real datasets.
• Outperformed competing methods in terms of c-index on MNIST and Cancer dataset simulations.
• Achieved a c-index of 0.803±0.006 on the real TCGA dataset, surpassing naive aggregation methods.

Conclusions:

• DALAN effectively aggregates multiple histopathology images for comprehensive survival analysis.
• The model's attention and LSTM mechanisms enable robust patient-level survival prediction.
• This approach enhances the interpretability and accuracy of deep learning-based survival models in oncology.