Deep learning-based survival prediction for multiple cancer types using histopathology images.

Area of Science:

• Computational pathology
• Oncology
• Artificial intelligence in medicine

Background:

• Accurate prognostic information is crucial for cancer treatment and monitoring.
• Current methods like staging and molecular features offer insights but require improvement in risk stratification.
• Histopathology images contain rich data for predicting clinical outcomes.

Purpose of the Study:

• To develop and evaluate a deep learning system (DLS) for predicting disease-specific survival (DSS) across 10 cancer types.
• To assess the DLS's performance against established clinical variables.
• To explore the potential of AI in improving prognostic accuracy using histopathology.

Main Methods:

• Developed a weakly-supervised deep learning system using 9,086 slides from 3,664 cancer cases (The Cancer Genome Atlas).
• Evaluated the DLS on 3,009 slides from 1,216 cases, testing three survival loss functions.
• Utilized multivariable Cox regression analysis for performance assessment and comparison with baseline models.

Main Results:

• The DLS was significantly associated with disease-specific survival across 10 cancers (HR 1.58, p<0.0001) after adjusting for clinical factors.
• The DLS improved model performance (c-index) by 3.7% compared to a baseline model.
• The DLS effectively stratified patients within specific cancer stages, including Stage II and Stage III.

Conclusions:

• Deep learning applied to histopathology images can provide significant prognostic information across multiple cancer types.
• This AI-driven approach has the potential to enhance clinical decision-making and patient monitoring.
• Larger datasets are needed to refine model performance and reduce confidence interval width for future deep learning survival models.