DCTNet: A Fusion of Transformer and CNN for Advanced Multimodal Medical Image Segmentation

In the present investigation, we propose an advanced multimodal breast cancer segmentation framework, designated as DCTNet, which harnesses the synergistic capabilities of Convolutional Neural Networks (CNN) and Transformers. This innovative approach is designed to amalgamate and utilize diverse informational and feature-rich inputs from varied modalities, significantly refining the precision of automated delineation in breast cancer lesions. DCTNet incorporates dual CNN-based feature learning architectures to independently assimilate modality-specific features, concurrently minimizing cross-modality interference through an intricately structured encoder-decoder mechanism complemented by skip connections. Furthermore, we introduce a Transformer-based encoder dedicated to cross-modal shared learning, adept at extracting cohesive representations from multimodal inputs. These are seamlessly integrated with modality-specific features via a Cross-Modal Feature Fusion Module (CFM), thereby optimizing the feature representation through the CNN decoder pathway for superior segmentation outcomes. Rigorous experimental evaluations conducted on the DCI breast cancer dataset affirm DCTNet's capacity to either match or excel beyond the segmentation efficacy of prevailing advanced multimodal models. This exploration not only elucidates the efficacy and indispensability of integrating CNN with Transformer structures, the cross-modal feature fusion module, and multimodal contrastive loss in elevating the accuracy of breast cancer segmentation but also pioneers new directions for ensuing research in multimodal medical image analysis.