MedMambaLite-v2: Shared Selective Scan for Efficient Edge Medical Mamba

Area of Science:

• Artificial Intelligence in Medicine
• Medical Imaging Analysis
• Edge Computing for Healthcare

Background:

• AI-powered medical imaging supports clinical diagnosis, but State-Space Models (SSMs) like Mamba face deployment challenges due to computational complexity.
• Real-time and energy-efficient edge applications are limited by the sequential data flow and high resource demands of existing SSMs.

Purpose of the Study:

• To propose MedMambaLite-v2, a novel shared selective scan framework for accelerating SSMs on embedded edge platforms.
• To enhance computational efficiency and reduce model size for real-time medical image analysis at the edge.

Main Methods:

• Developed MedMambaLite-v2 with a channel-only transition mechanism and optimized Convolution (Conv) branch.
• Applied knowledge distillation for model compression, creating a smaller student model.
• Designed LiteSS2D hardware with parallelism and 8-bit quantization for efficient inference.

Main Results:

• MedMambaLite-v2 is 23× smaller than the MedMamba baseline with only 1.1% accuracy reduction across 10 MedMNIST datasets.
• The LiteSS2D hardware prototype shows a 9× latency reduction compared to a serial baseline.
• Up to 63% and 78% energy reduction per inference on NVIDIA Jetson Orin Nano and Raspberry Pi 5, respectively.

Conclusions:

• MedMambaLite-v2 offers a highly efficient solution for deploying advanced AI medical imaging models on edge devices.
• The framework enables real-time, energy-efficient medical image classification with minimal impact on diagnostic accuracy.
• This work facilitates the integration of powerful AI tools into clinical workflows at the point of care.