Interpretable survival prediction for colorectal cancer using deep learning.

Area of Science:

• Computational pathology
• Machine learning in oncology
• Cancer prognostics

Background:

• Interpretable prognostic features are challenging to derive from deep learning models in histopathology.
• Accurate prediction of disease-specific survival is crucial for stage II and III colorectal cancer patients.

Purpose of the Study:

• To develop and interpret a deep learning system (DLS) for predicting disease-specific survival in colorectal cancer.
• To identify human-interpretable features that explain the DLS predictions and uncover novel prognostic markers.

Main Methods:

• A deep learning system (DLS) was developed using 3652 colorectal cancer cases (27,300 slides).
• The DLS performance was evaluated on two independent validation datasets.
• Interpretable features were generated by clustering embeddings from a deep-learning-based image-similarity model.

Main Results:

• The DLS achieved a 5-year disease-specific survival AUC of 0.70 and 0.69 on validation sets, outperforming clinicopathologic features alone.
• Clinicopathologic features explained only 18% of the DLS score variance.
• Clustering-derived histologic features explained 73-80% of the DLS score variance and were highly prognostic.
• A specific clustering-derived feature, characterized by poorly differentiated tumor cell clusters near adipose tissue, was identified with high accuracy by annotators.

Conclusions:

• The developed DLS effectively predicts disease-specific survival in colorectal cancer.
• Interpretable histologic features derived from deep learning embeddings significantly enhance model interpretability and prognostic value.
• This approach can uncover novel, visually identifiable prognostic features for future validation.