Accelerating Medical Image Analysis on Edge Devices: A Study with MobileOne Architecture

Neural networks, particularly convolutional neural networks, have proven effective in medical image analysis tasks such as classification (for example, identifying possible diseases), segmentation (for example, separating healthy tissue from necrosed), and object detection (for example, localising tumours). However, the benefits of developing and using such models come at significant infrastructural costs. There is a long-standing hope that comparable performance may be achievable with models on commodity, portable "edge" devices. These devices can be deployed for real-time decision-making by local inferencing and can improve access to these technologies in underdeveloped regions. Edge models typically trade-off model performance to achieve faster inference times on low-power hardware in edge devices. However, recent advances in generic feature-extraction pipelines ("backbones") for edge devices suggest that it may be possible to reuse them to construct models that can perform faster inference and predict almost as accurately as larger server-based models. We demonstrate this here using case studies of detection and classification tasks for medical images. Specifically, we use the MobileOne backbone, which provides inference times of under 1s. We compare its performance on well-known classification and detection datasets (HAM10000 and HyperKvasir) against state-of-the-art (SOA) models in server and edge computing environments. Our results show that MobileOne outperforms other SOA models while running on both server and edge devices. We have verified that the edge-based model is deployable on a commercially available, mid-range Android phone with a commodity GPU and edge devices such as Nvidia Jetson Nano, Orin Nano and Raspberry Pi.