MLDAN:Multi-scale large kernel decomposition attention network for super-resolution of lung computed tomography images

Abstract

Computed Tomography(CT) is widely used in medical diagnostics, with high-resolution(HR) CT images providing essential details for disease diagnosis. However, factors like equipment limitations and motion artifacts can result in low-resolution(LR) CT images. To address this, researchers have explored image super-resolution(SR) techniques to restore HR images, but challenges remain, particularly with lung CT images. One key challenge is the diversity of lung structures, such as variations in the shape and size of the lung parenchyma, bronchi, blood vessels, and lesions, which standard convolutional neural networks(CNNs) struggle to recover. In this paper, a method based on multi-scale large kernel decomposition attention network(MLDAN) for lung CT image super-resolution is proposed. This method captures detailed features of lung structures through multi-scale large kernel decomposition attention(MLDA) mechanism. To further enhance irregular lung edge details, we designed an anatomy-aware dynamic fusion module(ADFM), which adapts to complex lung features using deformable convolutions. Additionally, preserving global information between lung structures and their surrounding tissues is crucial for disease diagnosis. To maintain this, we incorporated MLDA as a Token Mixer within the MetaFormer architecture to efficiently model global dependencies. Experimental results on the public COVID-CT dataset and the lung cancer dataset IQ-OTH/NCCD show that the proposed method outperforms state-of-the-art(SOTA) techniques in both objective metrics(PSNR/SSIM) and visual quality, while maintaining a reduced parameter count. This method has significant potential for application and provides support for the early diagnosis of lung diseases. Codes are available at https://github.com/tiantianbaipiao/MLDAN