Automatic detection and evaluation of AD based on AI deep learning 1.1 Dataset of
atopic dermatitis The dataset will be established from more than 10,000 clinical
images of AD patients for AI deep learning. Low-quality images will be excluded, and
the images contained the surrounding background will be cropped to include only the
AD lesions.
1.2 Labelling the clinical signs of skin lesions The labelling will be completed by
three certified dermatologists and three trained algorithm engineers. The
dermatologists will label the clinical signs including erythema, papulation, edema,
oozing, excoriation, lichenification, and dryness, and severity of each sign will be
evaluated and labelled on a four-point scale (0: none, 1: mild, 2: moderate, and 3:
severe). The result of each clinical sign in an image will be labelled as an example
of erythema-2, edema- 2, or oozing-3. After labelling the images, the dermatologists
and algorithm engineers verify the quality of the labelled images from both clinical
and labelling rules and cross-validate the accuracy of signs and severity. Images
that meet the requirements will be used for model training. During the labelling and
model training process, the relevant personnel will be unaware of all the private
patient information.
1.3 Model training The model training will be carried out after labelling of the
images.
An accurate and efficient semantic segmentation model will be trained to distinguish
abnormal skin lesion areas to identify all the clinical signs. A fast and accurate
pixel level skin segmentation model will be trained to determine the ratio of the
lesion area to the overall skin area. Besides, an efficient and practical method to
convert the segmented skin lesion area into real skin area units will be created to
achieve the accurate restoration of the true size as much as possible from the
distortion of the skin lesion because of the shooting distance, angle, or automatic
enhancement. The dataset will be divided for training, validation, and testing.
Images of 6,500 of 10,000 will be used in training and validation of the proposed
model, and images of the remaining 3,500 of 10,000 will be used for testing. After
training, combined with the different questionnaire items filled by patients, the
evaluative tools including EASI, SCORED, POEM, pp-NRS, and DLQI will be calculated
by the model.
Development of the AIADMS app The app will support the Android system and IOS
system, and it will be designed as two versions for both patients and clinicians
with the distinguished login entrance. The fundamental function of the app will
include "Push", "Reminder", "Upload", "Evaluation", and "Data management".
2.1 The "Push" function is designed to transmit information to patients and medical
staff. The pushed information could be received and displayed on the screen of the
mobile phone even if the app is not opened and the mobile phone is in the locked
screen state, and the users can set the time of receiving the pushed information by
themselves. For example, the predetermined time point for follow-up in clinics will
be presented as "You should come to see the doctor on next Monday, July 25, 2023".
The "Push" function can activate the use of app, increase the viscosity of users,
and drive the utilization of other functional modules.
2.2 The "Reminder" function is mainly used for reminding the patients of taking
medicine, uploading photos of skin lesions, self-evaluation, and scheduled
follow-up.
2.3 The "Upload" function is designed to help patients participate in the systemic
treatment. They can upload their photos of skin lesions, the description of
progresses of AD, or questionnaires.
2.4 The "Evaluation" function is developed to provide information for both patients
and medical staff. By uploading photos of skin lesions and filling in the different
questionnaire items, the app will automatically evaluate the severity of lesions and
calculate the EASI, POEM, PP-NRS, SCORAD, or DLQI scores. This function could help
patients know more about their situation of the disease, and take part in self-
evaluation and self-care as the T2T strategy recommended.
2.4 The "Data management" function is designed for medical staff to manage the
patients more conveniently and design the medical research. They can log in to the
app platform website to collect and export data, carry out statistical analysis and
big data mining. App itself can also make simple statistics and management of data.
For example, data such as EASI, POEM and PP-NRS score at the time points of before
treatment, 2 weeks, 4 weeks, 12 weeks and 6 months after treatment could be
automatically generated into statistical reports to presented in the form of
histograms or curves. App can also be further improved and updated to the new
version through the analysis of users' habits, and the function modules could be
optimized with the high frequency of use and the feedback from both medical staff
and patients.