GhostCell-Net: A Lightweight Deep Learning Framework with Mobile Deployment for Automated Classification of Arsenic-Exposed Cells

INTRODUCTION: Chronic arsenic exposure from contaminated drinking water remains a critical public health concern, particularly in regions that rely on private wells. Accurate assessment of the impact of arsenic on cellular morphology requires computational tools that are both effective and accessible outside centralised laboratories.

Objectives

This work aims to develop a lightweight yet accurate deep learning model for the automated classification of arsenic-exposed PC12 cells, and to deliver it as a complete, on-device diagnostic system rather than an offline research prototype.

Methods

We propose GhostCell-Net, which builds upon the MobileNetV2 backbone and introduces three plug-in enhancement modules: Multi-Scale Dilated Aggregation (MSDA) for capturing cell features across multiple spatial extents, Frequency-Channel Attention (FCA) for texture-aware channel recalibration using DCT-inspired frequency descriptors, and a Spatial Gating Unit (SGU) for efficient spatial attention via channel-split gating. These modules are combined through a Progressive Stage Fusion (PSF) mechanism that aggregates multi-granularity features for classification, and the network is trained with a two-phase schedule using differential learning rates. The trained model is exported to PyTorch Mobile and integrated into an Android application supporting real-time camera capture, batch image processing, multi-model switching and automated accuracy evaluation.

Results

On the Arsenic Image Dataset, GhostCell-Net reaches 93.162% test accuracy with approximately 2.7M parameters, improving on the attention-based CNN baseline by 4.542 percentage points while using roughly 8.7 times fewer parameters. In on-device evaluation it attains 86.27% accuracy in 1567 ms, against 84.31% in 73068 ms for the ResNet-CBAM baseline.

Conclusion

GhostCell-Net achieves competitive classification accuracy at a substantially reduced computational cost and demonstrates a practical path from model design to point-of-care mobile deployment for cellular image analysis