High-frequency enhancement and dual-model ensemble for language-guided COVID-19 pneumonia lesion segmentation

Abstract

Accurate segmentation of coronavirus disease 2019 pneumonia lesions in chest x-ray images is critical for disease quantification and clinical decision-making. However, existing methods suffer from blurred lesion boundaries, missed detection of tiny lesions, and limited generalization performance. To address these challenges, this paper proposes a language-guided multimodal medical image segmentation method integrating high-frequency enhancement and a dual-model ensemble. The method is developed along three dimensions: frequency-domain detail enhancement, spatial alignment optimization, and dual-model ensemble. First, a frequency-domain decomposition-based high-frequency enhancement module is designed to mitigate detail attenuation caused by encoder downsampling. Second, a two-dimensional spatial position encoding module is introduced to improve the spatial alignment between text descriptions and visual regions. Third, a dual-model ensemble architecture with differentiated frequency-domain masks is constructed to address the performance ceiling of individual models. Finally, experiments on two public datasets with different imaging modalities and data characteristics demonstrate that the proposed method consistently outperforms existing single-modal and multimodal baseline methods and achieves competitive performance comparable to state-of-the-art approaches, verifying the effectiveness and robustness of the proposed framework. On the QaTa-COV19 dataset, the Dice coefficient reaches 91.85% and the mean Intersection over Union (mIoU) reaches 84.98%. On the multi-institutional MosMedData+ dataset, the Dice coefficient reaches 78.48% and the mIoU reaches 64.66%. The proposed method provides effective technical support for clinical diagnosis and follow-up assessment of pneumonia.