Whole slide image based prognosis prediction in rectal cancer using unsupervised artificial intelligence

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Digital pathology

Background:

• Rectal cancer lacks effective prognostic markers.
• Computational pathology offers novel approaches for marker development.
• Predicting progression-free survival (PFS) is crucial for rectal cancer management.

Purpose of the Study:

• To construct a prognostic signature for rectal cancer PFS prediction.
• To utilize unsupervised artificial intelligence on whole slide images.
• To identify novel prognostic markers from tumor characteristics.

Main Methods:

• Developed a tumor detection model using transfer learning.
• Employed a convolutional autoencoder for deep feature extraction from tumor patches.
• Clustered tumor patches and calculated percentage of each cluster (PCF) for signature construction.
• Validated the signature using Cox regression and nested cross-validation.

Main Results:

• Tumor detection accuracy reached 99.3%.
• An optimal 9-cluster model was identified, forming 9 PCFs.
• The prognostic signature achieved a concordance index of 0.701 in the validation cohort.
• Kaplan-Meier curves demonstrated good risk stratification ability.

Conclusions:

• The developed prognostic signature effectively predicts PFS in rectal cancer.
• Bioinformatic analysis identified PCF-associated genes and enriched pathways.
• This AI-driven approach holds potential for clinical translation in rectal cancer care.