Abstract:

About one in five people will develop herpes zoster (HZ), a skin condition. If antiviral treatment is not started within 72 hours of diagnosis, HZ might cause chronic pain syndrome. Leveraging artificial intelligence for mobile HZ diagnosis can alleviate neuropathic pain and reduce the burden on clinicians, as well as the associated costs. However, visual corruptions such as motion blur and noise are common in clinical photos acquired by common mobile devices. The purpose of this research is to educate a deep neural network (DNN) that can accurately discriminate HZ from other skin conditions using user-submitted photos in a portable and robust manner. We propose a curriculum training approach, knowledge distillation from ensemble via curriculum training (KDE-CT), where a student network progressively learns from a more potent teacher network, to achieve robustness while preserving computational efficiency. We created a curated dataset of skin illnesses for HZ detection and tested the model's resilience against 75 different kinds of corruptions. Thirteen distinct DNN models were compared using both uncorrupted and tainted image data. The results of the experiments show that KDE-CT is superior than other approaches in terms of corruption robustness. To be used for mobile skin lesion analysis, we trained MobileNetV3-Small to obtain exceptional performance (93.5% overall accuracy, 67.6 mean corruption error) while using substantially less multiply-andaccumulate operations (549x fewer).

Description:

Biomedical image processing, Convolutional neural networks, Deep learning, dermatology.