High-Frequency Dataset of Facial Skin
Description
The Dataset includes High-Frequency Images (image sequences) of Facial Skin. The data were acquired using DUB SkinScanner75 with a 24MHz transducer. The original size of images was 3466x1386x3. The image sequences were collected during 4 sesions, with 44 patients. The data were acquired from 3 different locations on patient face. The folder names refer to the acquisition date. The file names are defined as: px_y_z, where x denotes patient ID, y refers to the facial location, and z is the image index in the acquired series. During the acquisition process, the HFUS series were registered. Each of them includes the image data suitable for further diagnosis (techncal - using CAD software, or medical) and noninformative images. The image usfulness was described by 3 experts, and the labels are included in data_desc.xls. The repository includes CNN models (VGG16, pretrained on ImageNet), which were trained to divide the dataset into this 2 groups (informative and noninformative). Four different models were considered: 1 - vgg16_net_tl_64_1_w25_ALL_exp_aug0.mat 2 - vgg16_net_tl_64_1_w25_ALL_normal_aug0.mat 3 - vgg16_net_tl_64_1_w25_ALL_strict_aug0.mat 4 - vgg16_net_tl_64_1_w25_ALL_strict_aug0_v1.mat 1 - first model, from 5-fold, trained using labels provided by Expert 1, 2 - first model, from 5-fold, trained using labels provided by Expert 2, 3 - third model, from 5-fold, trained using labels provided by Expert 3. 4 - first model, from 5-fold, trained using labels, which were the same for all the experts (the expert agree). Due to the limited space, the image data are provided in the size of 224x224x3. This dataset was used to train all the CNN models. If you are interested in the image data in original size, please contect the corresopnding author.