Performance Evaluation of Embedded Image Classification Models Using Edge Impulse for Application on Medical Images.

Area of Science:

• Embedded Systems
• Machine Learning
• Medical Imaging

Background:

• Portable medical diagnostic systems require efficient on-device processing.
• Edge machine learning (ML) offers a solution for resource-constrained environments.
• Evaluating ML algorithms on microcontrollers is crucial for practical implementation.

Purpose of the Study:

• To explore the application of edge ML in portable medical image diagnostics.
• To evaluate the performance of MobileNetVI and a three-layer CNN on a microcontroller.
• To assess memory footprint, accuracy, inference time, and power consumption.

Main Methods:

• Two ML algorithms (MobileNetVI, CNN) were trained using the Edge Impulse platform.
• Models were deployed on a Nordic nrf52840 microcontroller in 32-bit float and 8-bit integer formats.
• Inference was evaluated using Edge Impulse EON compiler and TensorFlow Lite (TFLite) interpreter.

Main Results:

• Performance metrics including memory footprint, classification accuracy, inference time, and power consumption were recorded.
• Both 32-bit float and 8-bit integer representations were tested.
• Comparisons were made between EON compiler and TFLite interpreter.

Conclusions:

• The study demonstrates the feasibility of deploying ML models for medical image analysis on edge devices.
• Performance trade-offs between different representations and inference engines were analyzed.
• Findings provide insights for developing efficient portable medical diagnostic systems.