数据集获取

git clone https://github.com/linjiemu/MMXU.git

curl -L -o repo.zip https://github.com/linjiemu/MMXU/archive/refs/heads/main.zip
unzip repo.zip

源站 README 摘录（使用方式）

Usage

• File description

  • MMXU-test.jsonl: the MMXU test split, containing 3 k examples.
  • MMXU-dev.jsonl: the MMXU development/training split, containing 118 k examples. Due to PhysioNet restrictions, we cannot release the dev data publicly. If you would like to contact us further regarding this, please feel free to email me at linjiemu@sjtu.edu.cn.

• Data Description

  Each entry in the MMXU-dev.jsonl files contains the following fields:

  • id: A unique numerical identifier for each data entry.
  • content_type: The category of the question, often indicating the type of change being assessed.
  • related_region_name: The specific anatomical area of interest in the medical image.
  • question: The specific question that needs to be answered by comparing the current and prior medical images/reports.
  • options: A dictionary of possible multiple-choice answers to the question.
  • answer: The key corresponding to the correct option.
  • explanation: A text-based justification for why the answer is correct, often referencing the medical reports.
  • cur_image_path: The file path to the most recent medical image.
  • cur_report: The complete radiologist’'s report corresponding to the cur_image_path.
  • cur_region_report: A specific snippet from the full report that describes the findings in the related_region_name.
  • cur_image_bboxs: A dictionary containing the bounding box coordinates ([x_min, y_min, x_max, y_max]) for anatomical regions in the current image.
  • prior_image_path: A list containing file paths to one or more previous images used for comparison.
  • prior_report: The complete radiologist’'s report for the prior image(s).
  • prior_image_bboxs: Bounding box coordinates for regions in the prior image. It can be empty if the region was not annotated previously.

• Images

  • Please apply for and download the MIMIC‑CXR‑JPG dataset from https://physionet.org/content/mimic-cxr-jpg/2.1.0/.
  • After downloading, either create a symbolic link from this project’s files directory to the files directory of the MIMIC‑CXR‑JPG dataset, or adjust the image paths accordingly.

• Report information

  • Our dataset is built on the MIMIC‑CXR‑JPG dataset, which is a restricted‑access resource. This means we cannot release the as

数据加载示例（图像类）

from PIL import Image
import glob, os

files = (glob.glob(os.path.join(path, "**", "*.png"), recursive=True)
       + glob.glob(os.path.join(path, "**", "*.jpg"), recursive=True)
       + glob.glob(os.path.join(path, "**", "*.tif"), recursive=True))
print("图像文件数:", len(files))
img = Image.open(files[0]); print("尺寸/模式:", img.size, img.mode)

# torchvision Dataset 方式:
# from torchvision import datasets
# ds = datasets.ImageFolder(path)  # 要求 子目录=类别

目录组织与标注格式以源站说明和下载后实际文件为准。

完整仓库：github.com/linjiemu/MMXU