---
license: apache-2.0
tags:
- Report Generation
- Whole Slide Image
- Pathological Reports
- In-context Learning
- Lung Adenocarcinoma
- Multi-center
size_categories:
- 100B<n<1T
extra_gated_fields:
  Full name: text
  Affiliation: text
  Type of affiliation:
    type: select
    options:
    - Academia
    - Industry
    - label: Other
      value: other
  Official email: text
  Please explain your intended research use: text
  I agree to use this dataset for non-commercial use ONLY: checkbox
---

# Dataset Name
MMF (Multicenter Microscopic Findings)

### Dataset Description
We collect a multicenter microscopic findings(MMF) dataset from pathology reports of lung adenocarcinoma, consisting of a total of 827 WSI-report pairs from three different medical centers. Unlike existing WSI report generation datasets, our MMF offers the following advantages: (1) It contains only content that can be observed in the given WSI, eliminating irrelevant information. (2) It is the first Chinese pathology report dataset that addresses the language limitations of existing medical MLLMs and enhances diversity in this research field. (3) It includes multiple medical centers and highlights multicenter discrepancies in pathology report generation.

### Dataset Structure
MMF
├── GDPH

│ ├── caption.txt

│ ├── caption_en.txt

│ ├── 07cbc764db8b4a42bbc3af68a4a7f5c5.dzi

│ ├── 07cbc764db8b4a42bbc3af68a4a7f5c5_files

│ ├── xxx.dzi

│ └── xxx_files

├── DHSMU

│ ├── caption.txt

│ ├── caption_en.txt

│ ├── xxx.dzi

│ └── xxx_files

└── ZJHSMU

  ├── caption.txt
  
  ├── caption_en.txt
  
  ├── xxx.dzi
  
  └── xxx_files

caption.txt is the raw Chinese caption.

caption_en.txt is the translated English caption.

The WSI format is .dzi convert from .svs from libvips, containing a xxx.dzi file and tiles in xxx_files.

Larger number of sub-folder in xxx_files indicates larger magnifications (the largest is 40x, the second largest is 20x, and so on).

The dzi file can be processed directly by the MLLM, while the raw svs file is too large to load.


### Supported Tasks and Leaderboards
Report Generation.

In-context learning for WSI.

WSI caption.

Pathological pretraining.


### Dataset Details

The MMF consists of three sub-datasets, divided by the name of medical centers, namely GDPH (GuangDong Provincial people’s Hospital), DHSMU (Dongguan Hospital affiliated to Southern Medical University) and ZJHSMU (ZhuJiang Hospital of Southern Medical University). Specifically, GDPH and ZJHSMU are scanned by Aperio-SS12181, where the MPP (microns per pixel) is 0.264168 at the max magnification of 40x. DHSMU is scanned by Aperio-KFBIO at 0.482769 MPP (max 20x magnification). 642 out of 827 WSIs are sourced from GDPH, which is used as the training set, validation set, and internal test set. While DHSMU (92 cases) and ZJHSMU (93 cases) are only used as external test sets to evaluate the report generation performance across different report types. All the collected cases belong to lung adenocarcinoma, where the reports for the same disease show large differences in the content and style, showing the difficulty of this task. Our dataset. These cases range from Jan. 2009 to Sep. 2022, and all cases come from different patients to avoid data leaks. The age of patients ranges from 26 to 85 years, with a mean age of around 59 years. Besides, female cases take the major proportion for GDPH and DHSMU. In addition to the original Chinese report, we also provide an English version via the Google translation to benefit a broader international research community. Only de-identified microscopic findings and WSI were used in this dataset for research purposes.