DermXDB - a dermatological diagnosis explainability dataset

Repository for the DermXDB dataset.

Data structure

The data is split into two folders: annotations and masks. All dermatologist annotation are available in the annotations folder. Each JSON file is named dermatologist_image-name.json, and contains labelling information structured as:

{
  "dataset": "Explainability dataset",
  "image": {
      // Image metadata
    },
  "annotations": [
    {
      "name": "Characteristic name",
      "bounding_box": {
        // Height, width, and (x, y) of the upper left corner of the bounding box
        // Appears only if the characteristic was localisable
      },
      "complex_polygon": {
        // Array of paths to trace each mask
      },
      "attributes": [
        // Array of additional terms, if the characteristic is a basic term
      ]
    },
    {
      "name": "Diagnosis or non-localisable characteristic",
      "tag":{}
    }
  ]
}

For ease of use, binary masks were added in the masks folder. The masks/instance_masks_annotations.csv file creates the relationship between the image_id, the mask_id as found in masks/masks, and the characteristic name. Masks are stored as black and white PNG images, where black represents the background and white the selected area.

Most images have been renamed during labelling. Their original filename and source dataset can be found in metadata.csv. This file contains the original filename as found in the original dataset, the source dataset, the diagnosis associated with the image, and the filename under which the image can be found in DermX.

Getting access to the original images

The authors have not yet received the right to redistribute the SD and DermNetNZ images. To access the original images, please contact the owners of the SD-260 dataset and DermNetNZ dataset. Access to this data is only allowed for research purposes.

Referencing this work

If you use this repository in your research, please cite the work as:

@article{jalaboi2023dermx,
  title={DermX: An end-to-end framework for explainable automated dermatological diagnosis},
  author={Jalaboi, Raluca and Faye, Frederik and Orbes-Arteaga, Mauricio and J{\o}rgensen, Dan and Winther, Ole and Galimzianova, Alfiia},
  journal={Medical Image Analysis},
  volume={83},
  pages={102647},
  year={2023},
  publisher={Elsevier}
}