Detection of Helicobacter pylori Infection in Histopathological Gastric Biopsies Using Deep Learning Models.

Area of Science:

• Digital Pathology
• Computational Pathology
• Medical Image Analysis

Background:

• Traditional diagnosis of Helicobacter pylori (HP) gastritis relies on manual microscopic examination of H&E-stained gastric biopsies.
• Digital pathology introduces challenges like image resolution limitations and interobserver variability in HP detection.
• Deep convolutional neural networks (DCNNs) offer potential for automated and accurate HP identification in whole-slide images (WSIs).

Purpose of the Study:

• To evaluate the efficacy of DCNN and AutoML models for detecting HP infection in histopathological gastric biopsy samples.
• To compare the performance of various pretrained DCNN architectures (InceptionV3, Resnet50, VGG16) and AutoML approaches.
• To assess the diagnostic accuracy and reliability of automated methods in digital pathology for HP gastritis.

Main Methods:

• Development and validation of DCNN and AutoML models using a dataset of 100 H&E-stained gastric biopsy WSIs.
• Selection of 45,795 patches for model training and development.
• Utilizing immunohistochemistry for prior confirmation of HP infection in the dataset.
• Performance evaluation using metrics such as AUC, accuracy, recall, F1 score, and MCC.

Main Results:

• InceptionV3, Resnet50, and VGG16 achieved an Area Under the Curve (AUC) of 1.
• InceptionV3 demonstrated superior performance with 97% accuracy, 100% recall, 97% F1 score, and 93% MCC.
• AutoML models (BoostedNet, AutoKeras) exhibited performance below 85% across key metrics.
• External validation of InceptionV3 yielded a global accuracy of 78%.

Conclusions:

• DCNN models exhibit significant potential for diagnosing HP infection in gastric biopsies compared to AutoML approaches.
• The optimal model performance can vary across different pathology applications, necessitating a problem-specific approach.
• The adoption of DCNNs in digital pathology can enhance diagnostic accuracy, reduce variability, and streamline workflows for HP gastritis detection.