HPyloriNet: H. Pylori Detection in Gastric Biopsy Slides

HPyloriNet is a deep learning framework designed to automate the detection of Helicobacter pylori (H. pylori) in Histopathology Whole Slide Images (WSIs). Given that manual examination is time-consuming and labor-intensive, this project aims to aid pathologists by highlighting areas of bacterial presence using Computer Vision techniques.
The system implements and evaluates two distinct approaches: Classification and Object Detection.

Code Availability & Privacy Notice

Please Note: Due to privacy regulations and agreements with our clinical partner LABOKLIN, the training code and the full dataset cannot be made publicly available.
This repository contains only the inference pipeline, designed to run predictions on new or sample histology slides. It demonstrates the region extraction, segmentation, and detection logic described in our research.

Project Overview

• Goal: Improve the efficiency of Histopathology examination for H. pylori.
• Data Source: 22 H&E-stained veterinary gastric biopsy slides provided by LABOKLIN.
• Key Architectures: ResNet-101 (Classification), YOLO11x (Detection), NuClick (Segmentation).

Methodologies

This repository contains inference code for the two detection strategies described in the associated paper.

1. Classification Approach

This method focuses on extracting candidate regions (gastric glands) and classifying them as positive or negative.

• Region Extraction: Uses template matching with gastric gland templates to extract $256 \times 256$ candidate patches.
• Backbone: Utilizes pre-trained models (specifically ResNet-101) as feature extractors.
• Classifier: A Multilayer Perceptron (MLP) trained with Focal Loss to handle class imbalance.
• Output: Binary classification of patches.

2. Object Detection Approach

This method utilizes a "Sliding Window Patch Inference" algorithm to locate specific instances of bacteria.

• Inference: Splits WSIs into $640 \times 640$ patches.
• Segmentation (ROI Refinement): Uses a custom-trained NuClick model (from MONAI) to segment gastric glands and filter irrelevant tissue.
• Detector: YOLO11x is used for the final detection of H. pylori instances.
• Metric: Optimized using Distance IoU (DIoU).

Dataset & Preprocessing

The research was conducted on 22 WSIs at 40x magnification. The following preprocessing logic is embedded in the inference pipeline:

1. Annotation: Ground truth regions provided by experts, refined via an active learning loop to catch missed positives.
2. Patch Extraction:
   • Classification: Template matching thresholds used to identify gastric glands.
   • Detection: Sliding window extraction.
3. Normalization: Patches normalized using ImageNet statistics.

Results

Performance was evaluated on a stratified test split of 7 slides.

Method	Recall	Precision	Metric Notes
Classification	56%	45%	Evaluated on extracted candidate regions.
Object Detection	52%	58%	Based on DIoU threshold of 0.15.

Acknowledgments

We sincerely thank LABOKLIN for providing the dataset and ground truth labeling instrumental to this research.

Citation

If you use this code or methodology, please cite:

@article{HPyloriNet2024,
  title={A Computer-Aided Detection System for Helicobacter Pylori in Gastric Biopsy Slides},
  author={Nayan Sharma, Omer Ahmed, Zeyi Lu},
  year={2024}
}