FrugalVision: Adaptive Medical Image Compression for Resource-Constrained IoT Environments

Abstract

e energy and hardware constraints of medical IoT devices limit the deployment of real-time medical imaging solutions in remote environments. This work presents FrugalVision, an integrated solution for the reliable transmission of medical images in degraded network environments. The system combines three innovations: a residual autoencoder with adjustable differentiable compression, a reinforcement learning agent (PPO)dynamically adapting compression parameters, and the MQTT protocol for data and metrics communication in a constrained network.

Validated on three Decathlon medical segmentation datasets (hippocampus, heart, brain tumors), the solution maintains exceptional image quality with average PSNRs of 37.63 dB, 33.05 dB, and 35.02 dB, respectively, consistently exceeding the clinical threshold of 28 dB.  Comparative analysis with state-of-art methods  demonstrates a significant superiority of frugalvision, with PSNR gains up to +8.76 dB compared to conventional methods.

Processing latency is 0.15 to 0.26 s per image, with adaptation decisions taking 0.6 to 2.3 ms. Modeled power consumption is 0.02 to 0.04 J on a standard architecture without dedicated hardware acceleration. The architecture enables real-time adaptation to network variations using a simulator based on real network traces from the 5GMeas dataset. This approach paves the way for the reliable deployment of medical imaging solutions in underserved areas.

DOI: 10.61416/ceai.v28i2.10112