A Deep Multi-Label Segmentation Network For Eosinophilic Esophagitis Whole Slide Biopsy Diagnostics

Area of Science:

• Digital Pathology
• Artificial Intelligence in Medicine
• Gastroenterology

Background:

• Eosinophilic esophagitis (EoE) is an allergic esophageal condition characterized by eosinophil infiltration.
• Current diagnosis relies on manual eosinophil counting in biopsies, which is labor-intensive and subjective.
• Accurate diagnosis and monitoring are crucial for effective EoE management.

Purpose of the Study:

• To develop and validate a machine learning (ML) pipeline for automated identification, quantification, and diagnosis of EoE from whole slide images (WSIs) of esophageal biopsies.
• To improve the efficiency and objectivity of EoE diagnostics.
• To establish a scalable platform for analyzing complex histopathological data.

Main Methods:

• A deep learning-based platform combining multi-label segmentation and dynamic convolution was developed to process WSIs.
• The model was trained and validated on a large cohort of 1066 WSIs from 400 patients across multiple institutions.
• Performance was evaluated using metrics such as mean intersection over union (mIoU), mean absolute error (MAE), accuracy, sensitivity, and specificity.

Main Results:

• The ML model achieved high performance in segmenting intact and non-intact eosinophils with an mIoU of 0.93.
• Local quantification of intact eosinophils demonstrated a low MAE of 0.611.
• The system achieved high diagnostic accuracy (94.75%), sensitivity (94.13%), and specificity (95.25%) for EoE disease activity detection.

Conclusions:

• The developed ML pipeline offers a state-of-the-art, automated solution for EoE diagnosis using WSIs.
• This approach addresses key challenges in digital pathology, including simultaneous detection of small features and efficient whole-slide analysis.
• The findings pave the way for automated EoE diagnostics and have potential applications in other histopathological conditions.