🧬 Automatic Helicobacter pylori Diagnosis

📌 Project Overview

Helicobacter pylori (H. pylori) is a bacterium strongly associated with gastritis and gastric cancer. Its diagnosis from immunohistochemistry (IHC) slides is traditionally performed by pathologists through manual visual inspection, a process that is time-consuming and prone to subjectivity, especially in low-density cases.

This project explores deep learning–based automated systems to assist in the detection of H. pylori from histopathology images. The goal is to provide a robust patient-level diagnosis by aggregating patch-level predictions extracted from whole-slide images.

Two different approaches are implemented and compared:

• System I: Anomaly Detection using Autoencoders
• System II: Patient-level Classification using a Gated Attention Mechanism

---

🎯 Challenge Objective

• Detect H. pylori presence from IHC whole-slide images
• Perform patch-level analysis and aggregate results into a patient-level diagnosis
• Compare different deep learning architectures and design choices
• Evaluate robustness using cross-validation and a holdout test set

---

🧠 Methodology

The proposed pipeline consists of three main stages:

1. Patch Extraction (Preprocessed)
   The dataset provides 256×256 image patches extracted from tissue borders, where H. pylori typically appears.

2. Patch-level Analysis

   • Anomaly detection via reconstruction error (System I)
   • Feature embedding and attention-based aggregation (System II)

3. Patient-level Diagnosis

   • Patch predictions are aggregated to produce a final binary diagnosis per patient

---

🔍 System I: Anomaly Detection via Autoencoders

System I formulates H. pylori detection as an anomaly detection problem.

Key Ideas

• Models are trained only on patches from H. pylori–negative patients
• The model learns the distribution of normal tissue
• Patches containing bacteria produce high reconstruction error
• Patch predictions are aggregated to diagnose each patient

Models

• Convolutional Autoencoder (AE)
• Variational Autoencoder (VAE)

Reconstruction Error Metrics

Three reconstruction error definitions were evaluated:

• Mean Squared Error (MSE)
• Mean Absolute Error on the Red Channel
• HSV-based Red Pixel Reconstruction Error (selected)

The HSV-based metric explicitly captures biologically relevant red staining and achieved the best performance across all experiments.

---

👁️ System II: Attention-Based Patient Classification

System II treats each patient as a bag of image patches and learns a single patient representation.

Pipeline

1. Extract patch embeddings using the encoder from System I
2. Apply a Gated Attention Mechanism to weight informative patches
3. Aggregate patches into a patient-level representation
4. Perform binary classification

This approach allows the model to focus on the most relevant regions while suppressing background noise.

---

📊 Dataset

This project uses the Quiron dataset, a collection of 245 whole-slide immunohistochemistry images of gastric tissue, each corresponding to a different patient.

Data Distribution

It consists of 245 patients, with each patient contributing a single whole-slide image (WSI). For the purposes of binary classification, the diagnostic labels are organized as follows:

• Negative Group (117 patients): Labeled as NEGATIVA.
• Positive Group (128 patients): Consolidates the BAIXA (Low) and ALTA (High) categories.

---

Image Preparation and Validation

To optimize the data for computational analysis, the following preprocessing steps were applied:

1. Patch Generation: WSIs were segmented into $256 \times 256$ pixel tiles, focusing primarily on the tissue's marginal areas.
2. Expert Annotation: A specific portion of the dataset features patches that were manually labeled by specialists.
3. Experimental Setup: This annotated subset is reserved for critical pipeline stages, including:
   • Cross-validation
   • Model validation
   • Selection of classification thresholds

📚 References

Cano, P., Caravaca, Á., Gil, D., & Musulen, E.
Diagnosis of Helicobacter pylori using autoencoders for the detection of anomalous staining patterns in immunohistochemistry images.
arXiv preprint, 2023.