Real-Time Deep Learning Framework for Dermatology Image Classification on Low-Power Embedded Devices

The utilization of deep learning (DL) in medical research and industry has witnessed substantial growth in recent years. A pivotal application involves employing DL for dermatology image classification tasks. However, the major challenges in such tasks, in terms of the scarcity and bias of high-quality labeled data, significantly hinder further advancement in this domain. Such data insufficiency gives rise to concerns regarding accuracy disparities across different demographic groups, which may ultimately lead to unfair outcomes. Additionally, complex and effective DL models are often unsuitable with low-power embedded devices, which hinders their usability in resource-limited environments. In this paper, we propose a DL framework to address these issues. Our major approach involves augmenting data with Gaussian white noise to generate synthetic data samples and employing knowledge distillation techniques to transfer valuable knowledge from a larger and more complex model to a smaller and more efficient counterpart. Through comprehensive experimentation on an open-access skin disease classification dataset, we demonstrate that our proposed framework significantly enhances the performance of DL models on low-power embedded devices, thereby optimizing the trade-offs among overall accuracy, fairness for different demographic groups, and inference latency on low-power embedded devices.
^1
1
Our code and experiments can be reproduced by utilizing the details provided in the Methodology section on image preprocessing and augmentation, model architecture, and training configurations.(https://github.com/yixinli19/Dermatology-image-classification). The MobileNet V3 is available at(https://pytorch.org/vision/stable/models/mobilenetv3.html). The Swin Transformer V2 is available at (http-s://pytorch.org/vision/stable/models/swin_transformer.html). The dermatology image dataset is available upon request from ESFair 2023. They can be reached at: https://esfair2023.github.io/ESFair/.