Spike-Based Attention Mechanisms for Enhanced Medical Image Segmentation

The accurate segmentation of medical images is essential for reliable Computer-Aided Diagnosis (CAD) and treatment planning. Although deep learning has advanced the segmentation accuracy, challenges, such as modality variability, low contrast, and imaging artifacts, still exist. This study introduces a novel integration of a spike-based attention module into a DenseNet-169 backbone, combining biological spiking dynamics with deep convolutional feature representations to improve the focus on salient image features. The architecture utilizes ResNet blocks in the encoder for hierarchical feature extraction and applies spike-based attention at both the bottleneck and decoder stages to enhance the performance and computational efficiency. Experiments on two benchmark datasets demonstrate a 4.2% relative gain in Dice coefficient and a 3.9% increase in Intersection-over-Union compared to state-of-the-art baselines, while preserving the inference latency under 50 ms on standard GPU hardware. An extensive ablation study confirms that the dual placement of the attention module and the choice of DenseNet-169 maximize both the accuracy and efficiency. These results highlight the potential of spike-based attention for advancing real-time, high-fidelity medical image segmentation, with implications for improved clinical workflows and edge device deployment.