In this page, a light field dataset of skin lesions (SKINL2), captured with a focused plenoptic camera with extended depth of field, is presented. This publicly available dataset is proposed as an enabling resource to contribute for:
To accomplish these objectives all light fields are provided in two formats, namely lenslet and matrix of views. A dermatoscopic image corresponding to each lesion is also provided, to allow comparison with other diagnostic methods.
The light fields capturing system is built around a Raytrix R42 focused plenoptic camera, with a 25 mm f/1.8 lens. As shown in Figure 1, the camera is placed inside a cylindrical, acrylic-built black housing. Using this setup the camera is always placed at the same distance from the object (skin lesion), ensuring optimized focusing conditions (at a distance d ≈ 6.1 mm) and the same magnification factor for different light field acquisitions.
The scene illumination is based on a ring of 5 neutral white 5050 LEDs, placed inside the tube (≈ 16.1 mm above the imaged object ), with the black coating preventing both, noticeable patterns associated with LED light reflection, and ambient light interference. The light intensity of the LEDs is controlled by a computer running dedicated software, thus allowing to adjust the illumination conditions according to exposure requirements.
The setup of the acquisition system includes the camera calibration, which was carried out according to the manufacturer’s documentation, using a 2.0 mm point pitch calibration target.
(A) (B) (C) (D)
The light fields of skin lesions were captured at a hospitalar facility (Centro Hospitalar de Leiria, Portugal), from patients previously screened by a medical doctor during dermatology clinical appointments. Other imaging acquisition procedures, such as dermatoscopy and standard photography, were also carried out in order to collect different types of images from the same lesion.
Given the variety of skin types and lesion tonalities, the light intensity of the LEDs in the acquisition setup was adjusted before capturing every light field, in order to prevent either over- or underexposure
The skin lesions were manually classified and organized based on clinician diagnosis according to ICD10 (International Classification of Diseases), and on histopathological analysis. Procedures related with image acquisition were evaluated and approved by a health ethics committee. The Procedure and purpose of the study were explained to all volunteers, that signed an informed consent.
The SKINL2 dataset comprises a total of 376 light fields acquired under similar conditions. The images were classified using eight categories, according to the type of skin lesion/ICD code:
A
B
C
D
E
F
GThe light fields in the dataset are representative of the diversity of observed pathologies, skin tonalities (phototype), lesion size and shape, amount of hair and whether a clinician’s manual mark exists on the skin. Each light field is identified in the dataset by means of a unique ID number.
The light field dataset is made available in two different formats: (i) a lenselet, with a resolution of 7716 × 5364 pixels (8-bit depth RGB components), and (ii) a matrix of 9 × 9 views, each with a resolution of 3858 × 2682 pixels, where each pixel is represented by 16-bit depth RGB components. This matrix of views was obtained from the light field, using Raytrix API [2], by adjusting the parameter VirtCamPinholeStd_ViewOffsetMM_g in function RxSetPar.
The views are horizontally and vertically spaced by 1mm, and the dimension of 9×9 was considered the ideal number of views for estimation of the depth maps, following the analysis in [3]. A regular dermatoscopic image of each lesion is also provided, with a resolution of 1920 × 1080 pixels, being each pixel represented by 8-bit depth RGB components.
Use the following reference when using images frm the this dataset.
[1] S. M. M. de Faria et al., "Light Field Image Dataset of Skin Lesions," 2019 41st Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Berlin, Germany, 2019, pp. 3905-3908. DOI: 10.1109/EMBC.2019.8856578