<p>Computed Tomography (CT) imaging has been widely employed as a critical tool in clinical diagnosis. In recent years, deep learning and image segmentation techniques have been increasingly applied to the identification of lesions in CT images, particularly for pulmonary inflammatory damage caused by various viruses. Certain diseases present honeycomb-like lesions, which are challenging to accurately identify due to their variable morphology and uncertain localization. Furthermore, the small scale of these honeycomb lesions makes precise segmentation even more difficult. To address these challenges, we leveraged prior knowledge of the radiological scale features of the lesions and proposed a preprocessing strategy tailored to honeycomb lung datasets. This strategy aims to eliminate redundant and non-informative data while facilitating instance segmentation. Building on this, we introduced a local lesion copy-paste data augmentation algorithm to ensure that lesions are accurately placed within the pulmonary region while increasing the data volume. To tackle the small-scale characteristics of honeycomb lung lesions, we designed the MSCA-Sp R-CNN model, an instance segmentation framework that integrates a multi-scale channel attention module and a dual sub-pixel convolution upsampling module. Experimental results on the CC-CCII and UESTC-COVID-19 part2 datasets demonstrate that the proposed strategy and model significantly improve the accuracy of honeycomb lung lesion segmentation. Moreover, the visualization analysis highlights the model’s superior ability to identify multi-scale lesions, particularly small lesions. Compared to other algorithms, the proposed approach exhibits more competitive performance.</p>