Convolutional Neural Networks Can Determine AD Severity

March 20, 2021
Allison Inserro

Deep learning algorithms to diagnose skin disease like atopic dermatitis could save time for providers while increasing accuracy, according to a pilot study.

A pilot study demonstrated that convolutional neural networks (CNNs), a type of deep learning algorithm, could score the severity of atopic dermatitis (AD) at a level comparable with dermatologists.

The authors noted that trained CNNs have been shown to work on other dermatological conditions, namely skin cancers, diabetic retinopathy, and onychomycosis, and they wanted to see if the same could be achieved for AD.

The most valid, consistent, and reliable measure for AD is the Eczema Area and Severity Index (EASI), which consists of 4 components scored from 0 to 3 according to severity (none, mild, moderate, and severe): erythema, induration/papulation, excoriation, and lichenification. The EASI calculation also includes measuring affected body surface area, scored on a scale from 0 to 6.

The researchers created a standard dataset of 8000 cropped clinical images showing AD from anonymized patient data from Seoul St. Mary’s Hospital. Of the 8000 cropped pictures, 5600 images (1400 images each for erythema, induration/papulation, excoriation and lichenification) were used to train the CNNs.

The dataset was created out of 24,852 clinical images of AD acquired from 2009 to 2017, with the lesion areas cropped to 224 by 224 pixels. Three dermatologists scored the severity from 0 to 3 for each EASI component, with the final score determined by consensus.

Five hundred images were assigned a severity score to create a dataset of 2000 scored images for each EASI component. The remaining 2400 images (600 images each for erythema, induration/papulation, excoriation, and lichenification) were used for validation. The CNN was also externally validated with 400 images each from a different hospital.

Adjustments were made to the brightness of the images to probe if the CNNs accurately measured scores as image brightness levels changed.

Each of the CNN models (ResNet V1, ResNet V2, GoogLeNet and VGG-Net) was trained with an image dataset, and the researchers determined which CNN was best for scoring each component of EASI.

Accuracy rates of the CNNs were 99.17% for erythema, 93.17% for papulation, 96.00% for excoriation, and 97.17% for lichenification compared with manual scoring by dermatologists.

CNNs trained with brightness-adjusted images achieved a high accuracy without the need to standardize camera settings. These results suggested that CNNs perform similarly to dermatologists for scoring erythema, papulation, excoriation, and lichenification severity.

“The use of CNNs may increase the accuracy of AD severity scoring, allowing an accurate treatment response for patients and improving rapport with patients to improve treatment compliance,” the researchers wrote.

There were some limitations to this process, they noted. CNNs could be improved with more clinical images per EASI component. In addition, the images used in this study were not diverse, as it was conducted in South Korea. With adjustments, however, the authors thought that the method could also work with other skin colors.

In addition, since EASI measurements include both the severity score of each component and the ratio of the patient's lesion area, further study is needed to determine how to recognize the area score automatically.

Reference

Bang CH, Yoon JW, Ryu JY, et al. Automated severity scoring of atopic dermatitis patients by a deep neural network. Sci Rep. Published online March 15, 2021. doi:10.1038/s41598