Comparative Study of Lightweight Segmentation Models for Medical Cell Imagery in Resource-Constrained Environments

Abstract

Abstract Precise medical cell image segmentation is an essential prerequisite for cell counting, pathological diagnosis, cancer grading and personalized treatment planning, yet mainstream deep learning segmentation models are hard to deploy in grassroots medical institutions and portable microscopic devices due to their massive parameters and high computational complexity. To address this critical issue, this study evaluated three representative convolutional neural network models for medical nuclear segmentation tasks under GPU-free, resourceconstrained conditions, using the 2018 Data Science Bowl dataset with diverse tissue types and imaging modalities. The models included the lightweight Half-UNet as the benchmark, transfer learning-based MobileNetV2-UNet with pre-trained weights, and real-time-oriented Fast- SCNN. Experimental results show that MobileNetV2- UNet achieves the optimal Dice coefficient of 0.9175 with relatively low resource consumption, significantly outperforming the benchmark; Fast-SCNN sacrifices some accuracy (Dice coefficient of 0.7504) but gains superior inference speed (215.8 FPS), which is 3.8 times that of the benchmark. This study provides empirical evidence and technical references for efficient automated medical image analysis on low-computing hardware, highlighting the necessity of flexibly selecting lightweight models according to different clinical demands.