An Efficient Cross-Task Consistency Network for Joint COVID-19 Lesion Segmentation and Pulmonary Burden Estimation in Chest CT

Abstract

Accurate segmentation of pulmonary lesions and quantitative estimation of disease burden from chest computed tomography (CT) are essential for objective assessment of COVID-19–related lung involvement. This paper presents an efficient cross-task consistency network for joint COVID-19 lesion segmentation and pulmonary burden estimation from chest CT. The proposed architecture combines a depthwise-separable residual encoder, multi-scale dilated context aggregation, a compact Transformer bottleneck, and a dedicated burden-regression branch to capture complementary local and global representations of pulmonary abnormalities. A differentiable cross-task consistency loss regularises the burden prediction against the lesion-to-lung ratio derived directly from the segmentation probability map, thereby establishing an explicit relationship between anatomical delineation and quantitative disease assessment. Experiments were conducted on the COVID-19 CT Segmentation Dataset with expert lung and infection annotations using a patient-independent train-validation-test protocol. On the held-out test cohort, the proposed method achieved a Dice similarity coefficient of 0.8746, an intersection-over-union of 0.7832, a precision of 0.8915, a recall of 0.8619, and a pulmonary burden mean absolute error of 0.0218. The network contains only 0.351 million trainable parameters, providing an effective balance between segmentation accuracy, quantitative burden estimation, and computational efficiency. The findings demonstrate that explicit cross-task anatomical consistency supports accurate lesion delineation and coherent pulmonary burden quantification within a compact model, establishing a practical foundation for efficient quantitative analysis of chest CT images and facilitating reproducible deployment across computationally constrained medical imaging environments.