REDSCAR

Description

The images of the dataset contain within their name the main metadata. The name structure of all images is where:

• image_id is a random string of 25 characters, including lowercase and uppercase letters and numbers, which uniquely identifies each image.
• INF is a boolean (represented with 0 or 1), representing if the image has been classified with infection (1) or without infection (0). This classification has been carried out by health specialists.
• CAP is an integer, representing the phase of the clinical study in which the image was captured. Right now, because there have only been two phases of clinical validation, it can take the values 0 or 1.
• RES is a boolean (represented with 0 or 1), representing if the image has a valid resolution (1), more than 500x500 pixels, and (0) otherwise.

The fact that there are images with low resolution is due to the fact that the whole dataset has been captured using the in-camera patient's devices, in order to faithfully recreate real environments. To provide researchers with more information about the distribution of the images, we summarise in the table below some statistics according to their resolution and according to the presence of infection.

		Infection
		0	1
Valid  resolution	0	159 (40.25%)	5 (1.27%)
	1	207 (52.53%)	23 (5.84%)

Masks for wound and staple location

To provide more information about the spatial location of the wound and the staples in each wound, two masks have been constructed for each image. In addition, to facilitate some visualisation tasks, these masks have been drawn over the original images in green. Thus, in the main dataset, the images are distributed in the following folders:

• IMAGES, containing the original images.
• GT_WOUND_MASK, containing the masks (white over black background) with the exact position of the wound. All images are binary.
• GT_STAPLES_MASK, containing the masks (white over black background) with the exact position of the staples. All images are binary.
• GT_WOUND_COLORMASK, containing the masks with the exact position of the wound drawn over the original images.
• GT_STAPLES_COLORMASK, containing the masks with the exact position of the staples drawn over the original images.

In all folders the images have the same name. In this way, the type of image is differentiated according to the folder where it is contained.

Samples

To illustrate an example of Redscar Dataset image, the different images associated to image 0l9UkdcFPV5otARFN8qus6yBF_infection=0_capture=1_resolution=1 are shown below.

Subsets

The Redscar Dataset contains, apart from all the images described above organised in different directories, a subset already prepared for machine learning tasks. In the SUBSETS folder there are two directories, which are:

• MACHINE_LEARNING_DATASET, which contains two folders, train and test, containing a stratified training and test partition, in the proportion 70%-30%, respectively, of all images. In this framework, stratified means that the proportion is preserved in both subsets for the different classes (presence of infection and valid resolution).
• MACHINE_LEARNING_DATASET-resolution=1, which contains two folders, train and test, containing a stratified training and test partition, in the proportion 70%-30%, respectively, of all images that have a valid resolution. In this framework, stratified means that the proportion is preserved in both subsets for the class of presence of infection.