TransFocalCAD: A Hybrid Transformer-CNN Architecture with Multi-Scale Attention for Enhanced Medical Image Detection and Segmentation

Abstract

Accurate detection and segmentation of medical images are critical for early diagnosis and treatment planning. Traditional convolutional neural networks (CNNs) offer strong local feature extraction capabilities, while transformers have demonstrated remarkable global context modeling. In this study, we propose TransFocalCAD, a novel hybrid architecture that integrates the strengths of both CNNs and transformers, enhanced with a multi-scale attention mechanism. The proposed model leverages CNN-based modules for capturing fine-grained local features and transformer encoders for understanding long-range dependencies across spatial dimensions. The multi-scale attention framework dynamically refines feature maps at various resolutions, thereby improving contextual awareness and boundary delineation. Extensive experiments conducted on benchmark medical imaging datasets, including CT and MRI scans, demonstrate that TransFocalCAD significantly outperforms existing state-of-the-art methods in terms of segmentation accuracy, detection precision, and computational efficiency. The results underscore the potential of hybrid architectures with multi-scale attention in advancing the performance of computer-aided diagnosis (CAD) systems.