Image classification and auxiliary diagnosis system for hyperpigmented skin diseases based on deep learning.

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Hyperpigmented skin conditions, including melasma (ML), nevus fusco-caeruleus zygomaticus (NZ), freckles (FC), cafe-au-lait spots (CS), nevus of Ota (NO), and lentigo simplex (LS), are common dermatological concerns.
• Accurate diagnosis of these conditions is crucial for effective treatment, yet objective classification methods are needed.
• Deep learning (DL) offers potential for analyzing medical images and assisting in objective classification of skin diseases.

Purpose of the Study:

• To evaluate and identify the optimal deep learning algorithm for the auxiliary diagnosis of common hyperpigmented skin diseases.
• To compare the performance of ten different pretrained deep learning models for classifying images of six hyperpigmented skin conditions.
• To determine the most suitable DL algorithm for developing an online clinical diagnosis system for hyperpigmented skin.

Main Methods:

• Ten pretrained deep learning models (VGG-19, GoogLeNet, InceptionV3, ResNet50V2, ResNet101V2, ResNet152V2, InceptionResNetV2, DesseNet201, MobileNet, NASNetMobile) were utilized.
• These models were trained and tested on a dataset of images representing six common hyperpigmented skin diseases.
• Performance was evaluated using accuracy and area under the curve (AUC) metrics.

Main Results:

• MobileNet demonstrated a high training accuracy of 99.52% and an AUC of 0.93.
• While test set accuracies varied across models, MobileNet showed competitive performance.
• Training times and parameter counts differed significantly among the evaluated algorithms.

Conclusions:

• MobileNet emerged as a promising deep learning algorithm for the auxiliary diagnosis of hyperpigmented skin conditions.
• Its performance metrics suggest strong potential for clinical application in an online diagnostic system.
• Further development and validation are recommended to integrate MobileNet into clinical practice for improved diagnostic support.