SydneyMTL - Gastritis Sydney System Golden Dataset

Data Description: Consensus-Based Hibou-L Embeddings for Gastritis Research (Updated Sydney System)

Overview & Objective: The Updated Sydney System (USS) for chronic gastritis often suffers from subjective diagnostic thresholds. This dataset, curated via pathologist consensus, provides a high-fidelity "Golden Dataset" to support the development of standardized, objective AI models for gastritis grading.

Data Content: The original gastric biopsy specimens were collected from Seegene Medical Foundation, Seoul, South Korea (https://pr.seegenemedical.com/). This dataset contains 366 feature embeddings of the gastric biopsy WSI images extracted using the Hibou-L foundational model. It represents the morphological signatures of the gastric mucosa across five key attributes: H. pylori, Neutrophil activity, Mononuclear cells, Glandular atrophy, and Intestinal metaplasia. The samples were curated using a stratified joint-distribution strategy to ensure a balanced representation of all severity grades (0–3).

Structure & Interpretation:
1. HDF5 Files (.h5): Each file contains features (1024-D embeddings), coordinates, and addresses for patches within a Whole Slide Image (WSI).
2. labels_and_prediction.csv: Provides ground truth consensus grades and the predictions from the SydneyMTL model. Note that atrophy=4 indicates cases where grading was not applicable due to the absence of the muscularis mucosae, reflecting real-world clinical constraints. For model predictions, the values are formatted as: "label(confidence)"

Applications: This dataset is specifically designed to advance research and development in gastritis diagnostics.
It can be utilized to:
- Develop and validate automated gastritis grading systems based on the Sydney System.
- Benchmark Multi-Instance Learning (MIL) models for multi-attribute classification in gastric pathology.
- Analyze morphological feature representations of various gastritis phenotypes using foundational pathology encoders.