Deep learning systems detect dysplasia with human-like accuracy using histopathology and probe-based confocal laser...

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Probe-based confocal laser endomicroscopy (pCLE) aids real-time Barrett's esophagus (BE) dysplasia diagnosis but has low sensitivity.
• Histopathology, the gold standard, faces challenges due to inter-observer variability among pathologists.
• Improving diagnostic accuracy is crucial for effective BE screening and management.

Purpose of the Study:

• To develop and evaluate deep learning models for enhanced diagnostic accuracy of pCLE videos and biopsy images in BE.
• To compare the performance of patch-level and whole-slide image models for biopsy analysis.
• To identify diagnostically relevant tissue features using Gradient-weighted class activation maps (Grad-CAMs).

Main Methods:

• Deep learning models were trained on 1970 pCLE videos and 897,931 biopsy patches/387 whole-slide images.
• Models classified data into squamous, non-dysplastic BE, or dysplasia/cancer categories.
• Grad-CAMs were used to visualize model attention on relevant tissue structures.

Main Results:

• pCLE models achieved 90% overall accuracy and 71% sensitivity for dysplasia.
• Patch-level biopsy models reached 90% overall accuracy and 72% sensitivity for dysplasia.
• Whole-slide image biopsy models demonstrated 94% overall accuracy and 90% sensitivity for dysplasia.
• Grad-CAMs confirmed model focus on medically relevant tissue regions.

Conclusions:

• Deep learning models achieve high diagnostic accuracy for esophageal dysplasia detection via pCLE and histopathology.
• These AI approaches show potential to augment current screening protocols, improving accuracy and efficiency.
• The models' performance rivals human expert accuracy in similar studies.