ReCAN: a light-weight residual channel attention network with alternating skip connection for robust medical image classification

The high accuracy with computational efficiency is still a challenge to the medical image classification. This paper presents a lightweight convolutional neural network, ReCAN (Residual Channel Attention Network), which is effective at classifying medical images. The ResNet-18 backbone of ReCAN is truncated and combined with a new five-stage channel attention mechanism based on hierarchies of reduction ratios and alternating skip-connections. Assessed over seven MedMNIST collections of varied imaging modalities- such as pathology, dermatology, retinal imaging, and microscopy- ReCAN demonstrates competitive or superior performance on multiple datasets while maintaining parameter efficiency. On PathMNIST, the model achieves an accuracy of 95.64 percent with 11·47 M parameters and 1.8 GFLOPs, comparable to state-of-the-art methods while offering substantially lower computational cost. Statistical significance testing across multiple runs was not performed and is acknowledged as a limitation. Multi-stage attention architecture and strategic skip connection patterns are proven to be effective as validated in systematic ablation studies. It is established that appropriately designed channel attention processes within CNNs can be applied, and they can be subject to better performance than computationally expensive hybrid systems without affecting the generalization capability. ReCAN achieves an efficiency-precision ratio suitable for resource-constrained clinical settings, requiring only 1.8 GFLOPs and 11·47 M parameters. Inference time, throughput, and memory consumptison measurements were not collected and are acknowledged as limitations.