MRD-UNet: a compute-efficient medical image segmentation architecture with multi-attention and parallel dilated fusion
TL;DR
MRD-UNet provides a practical balance between segmentation accuracy and computational efficiency and outperforms baseline CNNs and performs comparably to heavier transformer-based models while using significantly fewer parameters.
Abstract
Medical image segmentation plays a critical role in computer-aided diagnosis by enabling precise delineation of abnormalities in medical images. Accurate segmentation of these images is essential for early diagnosis and treatment planning. However, conventional models such as U-Net face challenges including boundary uncertainty, variations in image quality and limited generalization across datasets. This study presents MRD-UNet, a U-Net based segmentation framework built on a channel-disjoint multi-path dilated fusion (MDF) block deployed uniformly across all encoder and decoder stages, a residual formulation within each MDF block that stabilizes gradient flow without introducing pooling branches, and a task-driven distribution of three functionally distinct attention modules. We evaluate the proposed model through extensive experiments on the CVC-ClinicDB dataset comprising 612 endoscopic images and the ISIC-2018 dataset containing 2,594 dermoscopic images. During data augmentation, we apply geometric transformations, elastic deformations, and brightness adjustments; training is performed using a composite loss that combines Dice loss and binary cross-entropy. The proposed method achieves a DSC of 0.9140 on CVC-ClinicDB and 0.9106 on ISIC-2018, yielding an improvement of approximately 2–3% over the baseline U-Net architecture. Ablation studies confirm that the channel-disjoint MDF design reduces the number of parameters relative to the baseline while improving sensitivity to small and low-contrast lesions, and that the targeted attention layout enhances boundary delineation. MRD-UNet provides a practical balance between segmentation accuracy and computational efficiency. It outperforms baseline CNNs and performs comparably to heavier transformer-based models while using significantly fewer parameters. Cross-dataset evaluations demonstrate its ability to generalize within the same imaging modality. Notably, the lightweight MRD-UNet-S variant maintains competitive accuracy, showing the architecture’s suitability for resource-constrained medical image segmentation tasks. Not applicable.