A lightweight deep learning model with channel attention for kidney cell classification from microscopy images

Accurate identification of renal cell types within tissue architecture is fundamental for understanding normal kidney physiology and detecting early pathological changes. While traditional histological examination is time-intensive and dependent on expert interpretation, deep learning based computational methods offer a scalable and reproducible alternative for large-scale cell classification. Existing general-purpose models are often over-parameterized and computationally inefficient when applied to resource-constrained settings. To address these limitations, this study introduces CytoECA-Net, a task-specific convolutional neural network optimized for classifying kidney cell types from fluorescence microscopy images. The proposed architecture follows a hierarchical five-stage design that employs depthwise separable convolutions to reduce redundant spatial filtering, combined with efficient channel attention and residual connections to capture discriminative intra-nuclear patterns. Evaluated on the TissueMNIST benchmark comprising approximately 236,000 images, CytoECA-Net achieves a classification accuracy of 75.56% and an AUC of 0.9564, with performance comparable to existing baseline architectures while using only 1.7 million parameters. It further demonstrates efficient computation with an inference time of 2.43 ms per image and low memory requirements. Additional evaluation on the KMC-RENAL histopathology dataset shows that the model achieves 97.76% accuracy and attains the highest performance among the evaluated models. Visual interpretability analysis using Grad-CAM confirms that the model focuses on biologically relevant nuclear structures. These results demonstrate that CytoECA-Net offers an effective balance of accuracy, efficiency, and interpretability for kidney cell classification, making it well suited for resource-limited biomedical and diagnostic environments.