Dual-Stream Dynamic Fusion Network for Medical Image Tumor Detection

Medical image tumor detection faces significant challenges in handling scale variations, ambiguous boundaries, and computational efficiency. To address these issues, this paper proposes a Dual-Stream Dynamic Fusion Network (DSDF-Net), which synergistically integrates convolutional neural networks (CNNs) and vision transformers through two innovative components. First, the Cross-Modal Feature Alignment (CMFA) module bridges the distribution gap between local CNN features and global transformer representations by leveraging adaptive instance normalization and learnable alignment coefficients. Second, the Adaptive Routing Attention Mechanism (ARAM) dynamically adjusts fusion weights based on lesion characteristics, prioritizing CNN features for small tumors and transformer features for diffuse lesions. Extensive experiments on BraTS 2021 (brain MRI), LiTS 2017 (liver CT), and a proprietary bladder cancer endoscopy dataset demonstrate state-of-the-art performance. DSDF-Net achieves an 89.7% Dice score, outperforming Swin-UNet by 2.4% while reducing parameters by 41%. Notably, it excels in detecting sub-5mm tumors with 76.8% recall (12.5% higher than TransFuse) and maintains real-time inference at 34.7ms/volume on NVIDIA T4 GPUs. Ablation studies validate that CMFA improves feature consistency by 42.7% in JS divergence reduction, and ARAM enhances small tumor detection by assigning 0.820.05 CNN weights for lesions <3mm. These advancements highlight DSDF-Net's potential for clinical deployment in early cancer diagnosis and intervention.