Research on Medical Image Analysis for Edge Devices Based on Lightweight Frameworks

Medical object detection underpins many computeraided diagnosis (CAD) workflows and remains central to clinical image analysis. In real applications, however, detection models must usually balance reliable accuracy against tight memory and computation budgets, especially on edge hardware. Although the YOLO family is widely adopted for real-time detection, its computational cost still limits deployment on embedded and resource-constrained devices. To address this problem, we propose YOLO-GCE, a lightweight framework that introduces Ghost modules to reduce backbone redundancy, a Cross-Scale Feature Fusion Module (CCFM) to strengthen semantic interaction in the neck, and an Efficient Upsampling Convolutional Block (EUCB) to suppress upsampling artifacts and improve smallobject detection. These components are designed to raise feature utilization without sacrificing inference efficiency, and the final model is further deployed on an RK3588s development board. Experiments on the BCCD and Br35H datasets show a 38.3% reduction in GFLOPs and a 50.5% reduction in parameters while maintaining strong detection performance. With only 1.49 million parameters, YOLO-GCE remains competitive with conventional baselines, supporting its use for real-time edge deployment in practical medical scenarios.