GEM-Net: a compact Ghost-ECA CNN with multi-stage feature fusion for multi-domain histopathology classification

Abstract Deep learning has achieved progress in medical image analysis, yet high computation of CNN and Transformer models limits deployment. This study introduces GEM-Net, a novel ultra-lightweight architecture for histopathological image classification. GEM-Net integrates three key innovations: (i) Ultra-Lightweight Ghost modules, which generate redundant feature maps inexpensively, (ii) a Micro-ECA attention mechanism, which adaptively calibrates channel responses using minimal parameters, and (iii) depthwise separable convolutions within compact residual Micro-Blocks, reducing computation while preserving representational power. Unlike prior lightweight models that apply efficiency modules independently, GEM-Net adopts an attention-regulated feature-economy design that tightly couples lightweight feature generation, capacity-aware channel recalibration, and compensatory multi-stage feature fusion. A knowledge distillation strategy further compresses the model, transferring knowledge from a 972.3K-parameter teacher to a 29.26K-parameter student with negligible accuracy loss. Extensive evaluations on five datasets–HER2-IHC (private), CRC100K, Kather-texture-2016, Chaoyang, and LC25000–demonstrate state-of-the-art or near state-of-the-art performance. GEM-Net achieves perfect accuracy on LC25000 and over 96% accuracy on HER2-IHC, CRC100K, and Kather, while outperforming most models on the challenging Chaoyang dataset. GEM-Net scales different variants for edge and server deployment. In addition, explainable AI techniques, including Grad-CAM++ and Integrated Gradients, provide interpretable visual and attribution-based explanations, enhancing transparency and clinical trust.