Hugging Face's logo

Datasets:
moukaii
/
Tuberculosis_Dataset

Tasks:
Feature Extraction
Text Classification
Modalities:
Image
Text
Formats:
parquet
Languages:
English
Size:
1K - 10K
Tags:
biomedical
imaging
computer vision
tuberculosis
multimodal
Libraries:
Datasets
pandas
License:
Dataset card Viewer Files Community
1
Tuberculosis_Dataset / README.md
moukaii's picture
moukaii
Update README.md
f5f8a56 verified
preview code
|
raw
history blame contribute delete
7.85 kB
metadata
language:
  - en
license: apache-2.0
size_categories:
  - 100M<n<1B
task_categories:
  - feature-extraction
  - text-classification
tags:
  - biomedical
  - imaging
  - computer vision
  - tuberculosis
  - multimodal
dataset_info:
  features:
    - name: case_id
      dtype: string
    - name: gender
      dtype: string
    - name: age
      dtype: int8
    - name: case_text
      dtype: string
    - name: keywords
      dtype: string
    - name: image_file
      dtype: image
    - name: caption
      dtype: string
  splits:
    - name: train
      num_bytes: 70088819.588
      num_examples: 6284
  download_size: 42809832
  dataset_size: 70088819.588
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Multimodal Dataset of Tuberculosis Patients including CT and Clinical Case Reports

Zhankai Ye
NetID: zy172

Dataset Summary

This dataset is curated from the original “The MultiCaRe Dataset” to focus on the chest tuberculosis patients. This is a multimodal dataset consisting of lung computed tomography (CT) imaging data and the clinical case records of tuberculosis patients, along with their case keywords, the captions of their CT images, patient_id, gender, and age information.

Dataset Sources

Supported Tasks:

This dataset can be utilized for:

  • Developing algorithms of the segmentation of chest CT images and the classification of tuberculosis positive or control.
  • Developing novel natural language processing (NLP) methods and unsupervised machine learning methods to extract medical terms from clinical notes.

Languages:

English

Data Structure and Instance:

The data will follow the structure below: {

  • "case_id": "PMC10129030_01",
  • "gender": "male",
  • "age": 62,
  • "case_text": "A 62-year-old man presented with acute dyspnea at rest, requiring high-flow…",
  • "keywords": "["dendriform pulmonary ossification", "lung transplant", "pulmonary fibrosis"]",
  • "pics_array": image
  • "Caption": "coronal. chest CT shows ground-glass and reticular opacities in the dependent…" }

Data Fields:

  • case_id (string): ID of the patient, created combining the PMC of the article plus a sequential number.
  • gender (string): Gender of the patient. It can be either Female, Male, Transgender or Unknown.
  • age (int): Age of the patient. Ages lower than 1 y.o. are assigned 0 as age.
  • case_text (string): Self-explanatory.
  • keywords (string): Keywords are taken from the keywords section that is sometimes available in the content of the article.
  • pics_array (int): image
  • Caption (string): Image caption.

Initial Data Collect and Preprocessing

  1. The original MultiCaRe Dataset, approximately 9GB in size, encompasses a diverse range of medical specialties including oncology, cardiology, surgery, and pathology. To create your tuberculosis-focused subset, the dataset undergoes a filtering process based on specific criteria:
  • Case Report Selection: The selection criterion for case reports is the presence of keywords such as 'tuberculosis' or 'tb'. This ensures that only reports relevant to tuberculosis are included.
  • Caption Filtering: The dataset is further refined by filtering captions that contain keywords like 'ct', 'lung', or 'chest'.
  • Image Labeling: Finally, the images are chosen based on the presence of labels 'ct' and 'lung'. This dual-label requirement ensures that the selected images are relevant to CT scans of the lungs, which are instrumental in detecting and assessing tuberculosis.
  • Through this meticulous filtering process, an initial tuberculosis dataset is compiled from the broader MultiCaRe Dataset. This dataset is messy, contains many diffferent files.
  1. To enhance the quality and relevance of the tuberculosis dataset, additional processing steps are implemented after in the Hugging Face python script after the initial filtration from the MultiCaRe Dataset:
  • Exclusion of Records with Missing Age Information.
  • Merge of data from difference files, including .csv, .JSON, and .jpg.

Social Impact

The multimodal dataset of tuberculosis patients, meticulously curated from the larger MultiCaRe Dataset, stands to have a significant social impact, particularly in the field of public health and medical research. Tuberculosis (TB) remains a major global health issue, especially in low- and middle-income countries, and the integration of CT imaging with clinical case reports in this dataset provides a rich resource for advanced diagnostic and treatment research. By facilitating the development of more precise algorithms for CT image segmentation and classification, as well as enhancing natural language processing (NLP) techniques for extracting medical terms from clinical notes, this dataset has the potential to improve the accuracy and efficiency of TB diagnosis.

Personal and Sensitive Information

Case reports are designed with the intention of being publicly accessible, and as a result, they deliberately omit any personal identifying details of the patients to ensure their privacy and confidentiality.

Bias, Risks, and Limitations

Bias

  1. Selection Bias: The original MultiCaRe Dataset was generated from 75,382 open access PubMed Central articles spanning the period from 1990 to 2023. Therefore, the random sampling of the cases from difference demographic groups cannot be guaranteed. The data may have bias as the collection process was not representative of the broader population. For example, the dataset may predominantly includes cases from a specific geographic location, age group, or socioeconomic status, and the findings may not apply to other groups.
  2. Technology Bias: Advanced imaging technologies might not be equally available in all settings, leading to a dataset that disproportionately represents patients from better-equipped facilities. This can skew the dataset towards conditions that are more likely to be diagnosed in such settings.
  3. Interpreter Bias: For the "case_text" and the "caption", variability in the expertise and experience of radiologists or clinicians interpreting the images can lead to differences in diagnosis or findings reported in the dataset.

Risks

  1. Privacy and Confidentiality Risks: Patient data, including case records and images, are highly sensitive. There's a risk of identifying individuals even if the data is properly anonymized.
  2. Data Integrity and Quality Risks: Inaccuracies, missing data, and inconsistencies within the dataset can compromise the validity of research findings or clinical decisions based on the data. This could lead to ineffective or harmful interventions.

Limitations

  • Data Quality:
  1. For textual data, certain patient records are missing key descriptive terms. Meanwhile, cases where imaging studies were not conducted lack both the images and their respective descriptive captions.
  2. Regarding images, a primary concern is also the incomplete nature of the dataset, as images do not accompany all patient records. Additionally, the image resolution varies, which can impede detailed examination. The inconsistency in image sizes and variations in the positioning of patient photographs may also pose challenges for consistent image analysis.

Citation 

@dataset{NievasOffidani2023MultiCaRe,
  author       = {Nievas Offidani, M. and Delrieux, C.},
  title        = {The MultiCaRe Dataset: A Multimodal Case Report Dataset with Clinical Cases, Labeled Images and Captions from Open Access PMC Articles},
  year         = {2023},
  version      = {1.0},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.10079370},
  url          = {https://doi.org/10.5281/zenodo.10079370},
}