SemilleroCV/BreastThermography

Dataset Card for Breast Thermography

This dataset contains breast thermography images, transformed from lossless radiometric JPEG images using the thermal parser tool. The calibration process leverages exiftool to generate a temperature text file from each image, with the final images stored in TIFF format for improved reproducibility and usability.

Dataset Details

Dataset Description

The original dataset was published on 5 February 2024 (Version 3) and is documented at Mendeley Data. It contains thermographic images of the female thorax area captured in a medical setting with a FLIR A300 camera. The dataset comprises three positions per patient (anterior, left oblique, and right oblique) and includes a variety of pathology results related to breast cancer research. Our contribution transforms the original JPEG images using the thermal parser repository to extract temperature data and converted to TIFF for compatibility with deep learning frameworks. This new format is intended to facilitate further analysis and experimentation. Additionally, clinical metadata has been used to generate descriptive text prompts via a custom Python function, using age, weight, height and protocol information to facilitate multimodal research.

• Curated by: Brayan Quintero, Miguel Pimiento, Guillermo Pinto, Julian León, Dana Villamizar.
• License: CC BY NC 3.0 (as per the original dataset).

Dataset Sources [optional]

• Repository: Mendeley Data Original Repository
• Paper [optional]: DOI: 10.1016/j.dib.2024.110503

Uses

Direct Use

This dataset is suitable for research in breast imaging, including but not limited to:

• Developing and testing image processing algorithms.
• Conducting experiments in passive breast cancer classification.

Out-of-Scope Use

• The dataset is not recommended for applications outside medical imaging and research.
• Misuse in any diagnostic context without proper clinical validation is strongly discouraged.
• Not intended for commercial use due to the CC BY NC 3.0 license.

Dataset Structure

The dataset is structured into training and testing splits. Each split contains TIFF images, generated text prompts and precomputed text embeddings (using GatorTron-Base) and segmentation masks. The images follow the original capture protocol and include metadata about the capture settings and patient position (anterior, left oblique, right oblique). Alongside descriptive sentences generated from the clinical metadata using a custom Python function (see code snippet here).

Dataset Creation

Curation Rationale

The dataset was curated to enhance the usability of the original thermographic images by converting them from JPEG to a TIFF format. This transformation facilitates reproducible research and enables easier integration with modern image processing workflows. In addition, the integration of descriptive clinical metadata allows the clinical context to be incorporated into image analysis workflows.

Who are the source data producers?

The original dataset was produced by a group of contributors including: Steve Rodriguez-Guerrero, Humberto Loaiza Correa, Andrés-David Restrepo-Girón, Luis Alberto Reyes, Luis Alberto Olave, Saul Diaz and Robinson Pacheco.

Personal and Sensitive Information

The dataset contains medical images and related metadata. While no explicit personal identifiers are provided, users should be aware that the images originate from a clinical setting and caution should be exercised regarding patient privacy and data sensitivity.

Bias, Risks, and Limitations

• The dataset reflects the clinical and demographic conditions of the patient cohort in Cali, Colombia, which may not generalize to other populations.
• The calibration process depends on the accuracy of the thermal parser and exiftool, and potential errors in these tools might affect the temperature readings.

Recommendations

Users should be aware of the inherent biases in the clinical setting of data collection and the technical limitations of the calibration process.

Additional Information

Citation Information

Dataset paper

@article{rodriguez2024dataset,
  title={Dataset of breast thermography images for the detection of benign and malignant masses},
  author={Rodriguez-Guerrero, Steve and Loaiza-Correa, Humberto and Restrepo-Gir{\'o}n, Andr{\'e}s-David and Reyes, Luis Alberto and Olave, Luis Alberto and Diaz, Saul and Pacheco, Robinson},
  journal={Data in Brief},
  volume={54},
  pages={110503},
  year={2024},
  publisher={Elsevier}
}

Our research paper

@INPROCEEDINGS{pintobreastcatt,
  author={Pinto, Guillermo and León, Julián and Quintero, Brayan and Villamizar, Dana and Rueda-Chacón, Hoover},
  booktitle={2025 IEEE Colombian Conference on Applications of Computational Intelligence (ColCACI)}, 
  title={Multimodal Vision-Language Transformer for Thermography Breast Cancer Classification}, 
  year={2025},
  volume={},
  number={},
  pages={1-6},
  keywords={Sensitivity;Translation;Mortality;Infrared imaging;Medical services;Metadata;Transformers;Breast cancer;Standards;Periodic structures;Breast cancer;deep learning;thermography;vision-language transformer;clinical metadata;cross-attention},
  doi={10.1109/ColCACI67437.2025.11230909}}