Researchers in the University of Southern Maine’s Department of Computer Science have developed a new artificial intelligence system that can identify cellular changes associated with arsenic exposure while being small enough to run directly on a mobile device.
The research, presented in the paper “GhostCell-Net: A Lightweight Deep Learning Framework with Mobile Deployment for Automated Classification of Arsenic-Exposed Cells,” was conducted by USM researchers Yunlong Li, Yuqi Song, and Xin Zhang, along with Chuxiong Wu of Southern Illinois University. The work combines deep learning, computer vision, and mobile computing to address a challenge in environmental health research: efficiently analyzing changes in cells exposed to arsenic.
Chronic exposure to arsenic through contaminated drinking water remains a significant public health concern. The researchers focus on PC12 cells, a model commonly used to study neuronal responses. Arsenic exposure can alter cellular morphology, including neurite growth and structural complexity. Traditionally, examining these changes in microscopy images can require substantial manual effort and can introduce variability between observers. The researchers sought to develop an automated approach that could analyze these images more efficiently.
The result is GhostCell-Net, a lightweight deep learning architecture built on the MobileNetV2 neural network. The researchers enhanced the model with four components designed to capture different characteristics of cellular images. These include Multi-Scale Dilated Aggregation, which captures features at different spatial scales; Frequency-Channel Attention, which focuses on texture-related information; Spatial Gating Unit, which identifies important spatial regions; and Progressive Stage Fusion, which combines information from different levels of the network.
One of the central goals of the project was efficiency. GhostCell-Net contains approximately 2.7 million parameters, compared with about 23.5 million in the ResNet-50-based attention model used as a previous benchmark. The researchers report that this represents an approximately 8.7-fold reduction in model size.
Despite its smaller size, GhostCell-Net performed strongly on the Arsenic Image Dataset. The dataset contains microscope images of PC12 cells exposed to five concentrations of sodium arsenite, ranging from an unexposed control group to 5.0 mM. GhostCell-Net achieved a test accuracy of 93.162%, compared with 88.620% for the attention-based CNN baseline.
The researchers also took the project beyond desktop-based model evaluation. They developed an Android application that runs the trained model directly on a mobile device using PyTorch Mobile. The application can acquire images through a camera, process batches of images, switch between different models, and report classification results.
The mobile results demonstrate the potential advantage of the lightweight approach. In an on-device evaluation involving 51 images, GhostCell-Net correctly classified 44 images with an accuracy of 86.27% and a total inference time of 1.567 seconds. The ResNet-CBAM model correctly classified 43 images, with an accuracy of 84.31%, but required approximately 73 seconds for inference.
This emphasis on mobile deployment is an important part of the research. Rather than focusing only on improving a model’s performance in a laboratory computing environment, the project considers what is required to make an AI system practical on resource-constrained hardware. The researchers envision applications in settings where access to conventional computing infrastructure may be limited, including field-based environmental health research and point-of-care settings.
The work also illustrates the potential for computer science research to contribute to interdisciplinary problems in environmental health and biomedical research. By combining computer vision with efficient neural network design, the researchers are exploring ways that AI can help researchers analyze biological responses to environmental exposure.
The authors emphasize that GhostCell-Net is an initial step rather than a finished clinical system. Future research will examine the model on larger, multi-site datasets and under different imaging conditions, including variations in microscopes, illumination, staining, image resolution, and sample preparation. The researchers also plan to investigate applications in other areas of toxicology and biomedical imaging.
GhostCell-Net demonstrates how advances in AI can move beyond increasingly large models toward systems designed with practical deployment in mind. By combining strong image-classification performance with a substantially smaller computational footprint, the project provides a promising foundation for future mobile applications in cellular and biomedical research.
