A Lightweight Model to Skin Disease Recognition

  • Duc-Quang Vu


Skin disease has become increasingly prevalent, emerging as
one of the most widespread conditions. It significantly affects human
health and even causes skin cancer and death. Therefore, there are many
methods have been proposed to solve this issue recently, especially deep
learning-based methods. However, these state-of-the-art methods seem
to only focus on how to achieve better performance and ignore the issue
of inference time. Specifically, deep learning-based methods usually build
very deep with a huge of model size and computational cost. As a result,
this becomes very difficult to deploy these models on devices with no
GPU support. In this study, we introduce a proficient and lightweight
model designed to address this issue, leveraging the Mobilenet architecture. Our experimental findings demonstrate that the suggested network
delivers comparable performance to contemporary cutting-edge techniques across diverse benchmark datasets, including HAM10000, International Skin Imaging Collaboration 2017, and International Skin Imaging
Collaboration 2019. Notably, our approach utilizes merely 0.2 million
parameters and 0.3 GFlops for image classification. This attribute holds
substantial importance for deploying the model on edge devices lacking
GPU support.


Ma T. Hong Thu1[0000−0001−6997−3757], To Huu Nguyen2[0009−0009−1780−456X],
Do Thanh Mai3[0009−0003−2036−2316], Trang Phung T.
Thu3[0000−0002−6801−1123], and An Dang4[0009−0000−2518−7180] Duc-Quang
1 Tan Trao University, Tuyen Quang, Vietnam
2 Thai Nguyen University of Information and Communication Technology, Thai
Nguyen, Vietnam
3 Thai Nguyen University, Thai Nguyen, Vietnam
4 Faculty of Computer Science, Phenikaa University
5 Thai Nguyen University of Education, Thai Nguyen, Vietnam