Self-contrastive weakly supervised learning framework for prognostic prediction using whole slide images.

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Digital pathology

Background:

• Automated prognostic prediction from histopathological images is challenging due to weak ground truth labels and the need to predict unobservable future events.
• Existing methods often struggle with the complexity and variability of histopathological data.

Purpose of the Study:

• To develop and validate a novel deep learning framework for automated prognostic prediction using histopathological images.
• To explore the significance of various regions of interest (ROIs) and employ diverse learning methods for real-world clinical applications.

Main Methods:

• A three-part framework combining convolutional neural network (CNN)-based tissue segmentation for ROI delineation, contrastive learning for feature extraction, and nested multiple instance learning (MIL) for classification.
• Initial validation on simulated data and a diagnostic task, followed by application to prognostic prediction in bladder cancer.

Main Results:

• The proposed framework was applied to bladder cancer prognostic prediction.
• The best models achieved an Area Under the Curve (AUC) of 0.721 for recurrence prediction and 0.678 for treatment outcome prediction on a private data cohort.

Conclusions:

• The developed deep learning framework demonstrates potential for automated prognostic prediction in histopathology.
• This research highlights initial findings on the limitations of current histopathological image analysis for predicting treatment outcomes and suggests avenues for future improvement.