Diverse Representation Knowledge Distillation for Efficient Edge AI Teledermatology in Skin Disease Diagnosis
Access to dermatological care in rural areas is limited due to a shortage of specialists. While AI-powered teledermatology offers a solution, it faces challenges from unreliable internet connectivity. Edge AI offers a promising approach that enables inferences locally on mobile devices. However, hig...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/11044323/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Access to dermatological care in rural areas is limited due to a shortage of specialists. While AI-powered teledermatology offers a solution, it faces challenges from unreliable internet connectivity. Edge AI offers a promising approach that enables inferences locally on mobile devices. However, high-performance AI models are often large, which makes them difficult to deploy on mobile devices. While recent research primarily concentrated on small models or cloud inference, this research addresses the underexplored application of knowledge distillation on mobile devices. By conducting applied research aiming at compressing the AI model while maintaining accuracy, this research proposes a novel adoption framework for skin disease inference using edge computing. The framework applies knowledge distillation to compress a high-performance teacher model into a small student model, followed by deployment on mobile devices for local inference. Key contributions include a novel framework for edge-based inference, evaluation of pre-trained models on the ISIC 2019 and Fitzpatrick17k-C datasets, and practical deployment on mobile devices. Model performance was evaluated using accuracy, precision, recall, and F1-score, while the prototype was measured using model size, compression ratio, and inference time. The experiments demonstrate that the scenario of distilling RegNetY32GF into MobileNetV2 results in the most efficient model, maintaining both accuracy and model size. The prototype evaluation shows the practicality of the proposed framework with a compression rate of 55.88 on the ISIC 2019 dataset, with inference time reduced by approximately 352 times. These results enable efficient inference on mobile devices with limited resources. |
|---|---|
| ISSN: | 2169-3536 |