CoDeSwiT: an efficient multi-stream hybrid architecture for global–local feature fusion via dynamic gating in medical image analysis

Medical image analysis plays a vital role in the early diagnosis of diseases. However, current deep learning approaches in this field have certain structural limitations. Although Convolutional Neural Networks (CNNs) demonstrate superior performance in extracting local features thanks to inductive bias, they fall short in capturing the global context. On the other hand, Vision Transformer (ViT) architectures are powerful in modeling global relationships; however, they have disadvantages such as loss of local detail, excessive data requirements, and high computational costs. CNN-ViT hybrid models typically use multi-branch structures that combine the outputs of independent and large backbones. This significantly increases the model's computational load and number of parameters, reducing efficiency and making it difficult to use in clinical applications. To overcome these limitations, this study proposes the original and efficient CoDeSwiT architecture. CoDeSwiT integrates local and global features into a low-cost architecture through its parallel arms designed on a single data stream and weighted fusion capability, unlike computationally intensive hybrid structures. The proposed architecture effectively reduces dimensions using the modern ConvNeXt stem for rich feature representation, while integrating the Swin Transformer encoder to model the global context hierarchically and computationally efficiently. The Depthwise Separable Convolution (DWS) and Squeeze-and-Excitation (SE) local attention branch, designed in parallel with the Transformer stream in the architecture, ensures the capture of fine texture details and local features. The outputs of these heterogeneous branches are fused through a learnable gating mechanism, unlike standard static fusion operations in the literature. The gating mechanism dynamically determines and weights the diagnostic importance of local or global information for each image sample, enabling the model to adaptively focus on critical findings. The effectiveness of the proposed architecture was evaluated on three comprehensive datasets with different modalities: the Alzheimer MRI dataset, the Chest X-Ray dataset, and the Tuberculosis dataset. Experimental results show that CoDeSwiT outperforms state-of-the-art (SOTA) transfer learning models based on CNN and ViT, achieving accuracy rates of 96.46% on the Alzheimer MRI dataset, 96.84% on the Chest X-Ray dataset, and 99.67% on the Tuberculosis dataset. CoDeSwiT, which contains approximately 28 million parameters, exhibits a computationally efficient structure compared to large-scale architectures in the literature. The model, which achieves a processing time of 2.21 ms per image, has been observed to have the potential to offer an effective solution for supporting fast and accurate diagnosis processes in clinical environments with limited hardware resources.