Constrained multiple instance learning for ulcerative colitis prediction using histological images.

Area of Science:

• Gastroenterology
• Computational Pathology
• Artificial Intelligence

Background:

• Ulcerative colitis (UC) is a chronic inflammatory bowel disease (IBD) with a relapsing-remitting course.
• Histologic remission (HR), assessed by neutrophil presence, is a key indicator of UC activity and treatment target.
• Current deep learning approaches have not focused on neutrophil detection in whole-slide images (WSIs) for UC identification.

Purpose of the Study:

• To develop and evaluate a novel deep learning framework for detecting UC activity using neutrophil detection in WSIs.
• To improve the accuracy of histologic remission assessment in UC.
• To establish a new computational method for analyzing WSIs in the context of UC.

Main Methods:

• A novel multiple instance learning (MIL) framework with location constraints was developed.
• The framework incorporates constraints on positive instances to leverage weakly supervised information.
• A weighted embedding strategy was employed to enhance the relevance of positive instances.

Main Results:

• The proposed MIL framework demonstrated significant improvements in WSI-level accuracy on the PICASSO-MIL dataset.
• The method achieved a 10% improvement in bag-level accuracy compared to prior MIL settings.
• Location information was shown to considerably enhance the performance of UC detection.

Conclusions:

• The developed model surpasses current state-of-the-art MIL methods for UC detection.
• The novel constraint approach offers a promising direction for histological analysis.
• The method has potential applications in other histological conditions with subtle morphological features.