Measurement of psoriasis affected area with an artificial neural network: Supplementary Material

Mendeley Supplementary Figure 1 shows the flow chart of dataset configuration. Dataset obtained from 149 patients who visited Seoul St. Mary’s hospital dermatology clinic from 2018 to 2020, mainly consisting of images for each area division used to calculate PASI score (i.e., trunk, upper extremities, lower extremities). Various psoriasis phenotypes (small plaque, large plaque, regressed lesion), non-psoriatic lesion, and artifacts (e.g., accessories, shadows) were included in the dataset. Excluded images (35 images) were facial images or out-of-focus images. To evenly distribute the patient severities in the training and test sets, the data set was first split into three groups (mild, moderate, severe) by PGA. This PGA-based split was organized as mild (grade 0-1), moderate (grade 2-3), and severe (grade 4). Random sampling was performed in each group in a 4:1 ratio. To evaluate the PASI area score, inappropriate images for applying the palm method were excluded and 297 images from the test set were used to measure the performance with three dermatologists. With respect to input image, automated psoriasis detector determines skin area and psoriatic lesion to calculate the percentage of affected area. Ground truth was obtained by manually annotating each lesion margin with agreement between three annotator dermatologists. Skin inference and prediction are the result of determined skin area and psoriatic lesion by automated psoriasis detector, respectively. Mendeley Supplementary Figure 2 shows the flowchart of the automated psoriasis detector. Cascade Mask R-CNN with a Swin Transformer small (Swin-S) backbone, the same model used in APD, was used to determine affected skin area with pretrained weights from the COCO dataset. Manual skin color-based thresholding was applied in a few cases with insufficiently unclothed images.