Light-M: An efficient lightweight medical image segmentation framework for resource-constrained IoMT.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Internet of Medical Things (IoMT)

Background:

• The integration of the Internet of Medical Things (IoMT) into healthcare offers enhanced monitoring but faces challenges due to resource limitations of standard IoMT devices and the demands of large AI models.
• Existing intelligent healthcare solutions struggle to balance sophisticated AI capabilities with the constraints of edge devices.

Purpose of the Study:

• To propose a novel Knowledge Distillation (KD)-based end-edge-cloud orchestrated architecture, Light-M, for deploying lightweight medical image segmentation models on resource-constrained IoMT devices.
• To address the paradox between the need for large AI models and the limitations of IoMT devices.

Main Methods:

• Developed Light-M, an IoMT architecture utilizing KD to train a lightweight student model on edge devices by imitating a powerful cloud-based teacher model.
• Implemented two KD strategies: Active Exploration and Passive Transfer (AEPT) for discovering new features and Self-Attention-based Inter-Class Feature Variation (AIFV) for enhancing class distinction.
• Ensured no additional computational overhead during deployment by confining KD strategies to the training phase.

Main Results:

• The Light-M architecture demonstrated improved student model learning representations through the combined KD strategies.
• Achieved superior performance compared to state-of-the-art methods on cardiac and real-scene medical image segmentation tasks.
• The distilled student model exhibited efficient inference, completing tasks in just 29.6 ms on an IoT device.

Conclusions:

• The proposed Light-M framework effectively deploys lightweight medical AI models on IoMT devices, overcoming resource constraints.
• The novel AEPT and AIFV KD strategies enhance model performance and efficiency for medical image segmentation.
• Light-M offers a viable solution for advanced intelligent healthcare services at the edge.