README.md

metadata

configs:
  - config_name: default
    data_files:
      - split: demo
        path: demo.tsv
      - split: classification
        path: '*classification*'
      - split: localization
        path: '*localization*'
      - split: segmentation
        path: '*segmentation*'
      - split: report_generation
        path: '*report*'

Dataset Description

Our OmniFM-Dr framework introduces a multi-task chest x-ray dataset, which is used for the joint training of disease classification, localization, segmentation, and report generation. This dataset comprises various publicly available datasets, such as MIMIC-CXR, VinDr-CXR, and ChestX-Det10. For each image, it can potentially contribute to multiple tasks, such as report generation and classification.

NOTE: Due to requirements related to data compliance and other regulations, the dataset is temporarily unavailable. However, for each task, we will provide a showcase of five samples.

Dataset Details

• MIMIC: contains more than 377,110 radiograph images from over 227,835 radiographic studies. Each radiograph is paired with lesion classification and associated radiology report. It is used for multi-label classification and report generation tasks.
• Padchest: includes 160,840 images obtained from 67,000 patients, covering six different position views. Different radiographic findings were labeled and used for the classification task in this study.
• CXR-AL14: is a large-scale dataset for chest X-ray image detection. It has more than 140,000 chest X-ray radiographs containing 253,844 bounding boxes in 14 chest abnormal object categories.
• VinDr-CXR: includes chest radiographs with annotations for the classification of 28 common chest diseases. The dataset contains 15,000 CXR scans in the training set. We select eight diseases from the dataset along with their corresponding bounding box for the disease localization task.
• ChestX-Det: consists of 3,578 images from NIH ChestXray14 for 13 common disease. We select seven diseases from the dataset along with bounding box for the disease localization task.
• CheXmask: contains 676,803 lung and heart segmentation masks of chest images from six publicly available databases: CANDID-PTX, ChestXray14, Chexpert, MIMIC-CXR, Padchest, and VinDr-CXR. We include 224,316 data for training and 10,000 data from ChestXray14 for downstream evaluation.
• SIIM: comes from the SIIM-ACR Pneumothorax Segmentation competition and contains 12,090 images, among which approximately 3,000 cases are positive for pneumothorax disease with masks.

Dataset Structure

• MIMIC:
  • MIMIC_classification_report-generation_xxx.tsv: is used for classification and report generation tasks. For each row, it contains id, report, "label1 && label2", subject_id, study_id, dicom_id.
  • MIMIC_classification-location_xxx.tsv: is used for location vqa task. For each row, it contains id, "label1,severity && label2, severity", subject_id, study_id, dicom_id.
  • MIMIC_classification-severity_xxx.tsv: is used for severity vqa task. For each row, it contains id, "label, location1 & location2", subject_id, study_id, dicom_id.
• Padchest:
  • Padchest_classification_xxx.tsv: is used for classification task. For each row, it contains id, "label1 && label2", subject_id, study_id, dicom_id.
• CXR-AL14:
  • CXR_AL14_localization_xxx.tsv: is used for locatization and classification tasks. For each row, it contains id, label, "x1,y1,x2,y2", image_id.
• VinDr-CXR:
  • VinDr_CXR_localization_xxx.tsv: is used for locatization and classification tasks. For each row, it contains id, label, "x1,y1,x2,y2", image_id.
• ChestX-Det:
  • ChestX_Det_localization_xxx.tsv: is used for locatization and classification tasks. For each row, it contains id, label, "x1,y1,x2,y2", image_id.
• CheXmask:
  • CheXmask_segmentation_xxx.tsv: is used for segmentation task. For each row, it contains id, label, "x1,y1,x2,y2, ..., x30, y30", subject_id, study_id, dicom_id.
• SIIM:
  • SIIM_segmentation_xxx.tsv: is used for segmentation task. For each row, it contains id, label, "x1,y1,x2,y2, ..., x30, y30", subject_id, study_id, dicom_id.

Dataset Use

• Please run data_prepare.py, which constructs a training batch for all tasks. Each row should contain the following: id, instruction, label, image_id, and task_type.