MC-DiffNet: Mask-Constrained and Cycle-Consistent Diffusion Network for Unpaired Ultrasound Image Enhancement

Abstract

Ultrasound imaging is a valuable tool in clinical diagnostics due to its real-time feedback and non-invasive nature. Despite these advantages, it often suffers from issues like low contrast, speckle noise, and blurred anatomical boundaries, which can compromise the accuracy of diagnoses. Although diffusion models have proven effective for image restoration, their use in enhancing unpaired ultrasound images is limited. Another limitation is lacking of structural constraints and region-specific guidance. To address these challenges, we introduce MC-DiffNet, a multi-task, diffusion-based enhancement framework designed specifically for clinical ultrasound images in unpaired settings. Furthermore, the use of lesion masks provides structural guidance that enables the model to focus on pathological characteristics in ultrasound images. The framework incorporates three principal modules: a Cycle-Consistency Path, a Mask-Constrained Module, and a Context-Aware Classification Path. The Cycle-Consistency Path ensures consistency in unpaired training through degradation-reconstruction consistency. The Mask-Constrained Module incorporates lesion-aware structural constraints into the Structural Similarity Index Measure (SSIM) loss, aiming to maintain anatomical accuracy. Meanwhile, the Context-Aware Classification Path guides the enhancement process using semantic-level features. Together, these modules enable MC-DiffNet to enhance diagnostically significant regions with precision while preserving anatomical integrity. Experiments conducted on public ultrasound datasets reveal that our method outperforms existing techniques in both peak signal-to-noise ratio and SSIM. Remarkably, when applied to real-world unpaired datasets, MC-DiffNet offers improved structural fidelity and visual clarity. Further evaluation suggests that the enhanced images generated by our framework are beneficial for subsequent segmentation tasks, underscoring the clinical relevance and robustness of our approach.