Authors:
Tess Watt、Christos Chrysoulas、Peter J Barclay
Paper:
https://arxiv.org/abs/2408.08215
Introduction
Artificial intelligence (AI) assistants have become invaluable in the healthcare domain, aiding in tasks such as diagnosis and patient monitoring. However, their reliance on internet connectivity limits their usability in remote and low-connectivity areas. According to a United Nations report, approximately 37% of the world’s population has never used the internet, highlighting the need for solutions that do not depend on cloud connectivity.
TinyML, which involves deploying machine learning (ML) algorithms on small, constrained devices, offers a promising solution. This study explores the feasibility of using tinyML to provide healthcare support in low-connectivity environments, focusing on the diagnosis of skin diseases. By training a model on 10,000 images of skin lesions and deploying it on a Raspberry Pi, the study aims to demonstrate the potential of tinyML in healthcare.
Related Work
Cloud Computing
Cloud computing has revolutionized data processing by providing on-demand access to a shared pool of resources. However, it faces challenges such as latency and high power consumption, especially in Internet of Things (IoT) scenarios. These limitations make cloud computing less suitable for hosting AI assistants in remote environments.
TinyML
TinyML and edge computing can address the connectivity issues associated with cloud computing by processing data locally on constrained devices. This approach reduces communication costs and energy consumption. Various devices, such as Raspberry Pis, UDOO BOLT, and Nvidia Jetson Nano, offer different capabilities in terms of memory and processing speed, making them suitable for different applications.
Image Classification
Image classification for skin lesion detection has been extensively studied using high-spec computers. For instance, Qin et al. achieved a 95.2% accuracy using a deep neural network on a high-performance computer. However, deploying such systems on low-spec devices without GPUs remains a challenge. Previous attempts, such as Ramlakhan & Shang’s mobile melanoma recognition system, achieved lower accuracy due to hardware limitations.
Ethical Considerations
The ethical use of AI in healthcare involves several key considerations:
- Data Collection and Storage: Healthcare data is highly sensitive, and its wrongful disclosure can lead to stigmatization or discrimination.
- Impact of Diagnoses: AI-generated diagnoses can lack transparency, potentially leading to incorrect treatments.
- Medical Verification: Software tools used for diagnosis must be approved as medical devices.
- Bias and Diversity: Ensuring diversity in training data and among developers can help mitigate biases in AI systems.
System Architecture
The study aimed to design a tinyML prototype for healthcare support, focusing on image classification for skin lesions. The Raspberry Pi 3 Model B was chosen for its affordability and sufficient processing capabilities. The HAM10000 dataset, consisting of 10,000 images of common pigmented skin lesions, was used to train the model.
The model was trained using the MobileNet-V2 architecture and then offloaded to the Raspberry Pi. Google’s MediaPipe API was used to bridge the image classification model with real-time video input. The system’s architecture and a sample of the dataset are shown below.
Evaluation and Results
The model achieved a test accuracy of 78% and a test loss of 1.08. The training and validation accuracy improved consistently with each epoch, indicating that the model did not suffer from overfitting. However, the accuracy and loss values suggest room for improvement.
A comparative assessment with existing models showed that the proposed model outperformed Ramlakhan & Shang’s system but lagged behind higher-spec systems. The table below summarizes the comparison.
Conclusion
This pilot study demonstrated the feasibility of deploying an image classification system for skin lesion detection on a constrained device. The Raspberry Pi 3 Model B was used to prototype the system, achieving a test accuracy of 78%. Future work will focus on comparing and tuning different algorithms, using model compression techniques, and exploring more powerful boards like the UDOO BOLT and Orange Pi 5. Integrating edge computing for federated learning could further enhance the system’s capabilities, paving the way for real-time detection of visually detectable diseases in remote and low-connectivity areas.
