LAUSR

Chest computed tomography (CT) has become an essential imaging modality in the early screening and clinical diagnosis of pulmonary diseases. Among its key tasks, abnormality detection has garnered increasing attention. Teacher–student (T–S) models trained solely on normal samples exhibit strong generalization capabilities, yet often suffer from overcorrecting anomalous regions during inference, potentially obscuring true lesions and compromising detection performance. In addition, conventional attention mechanisms—though effective in global modeling—are computationally expensive and pose scalability challenges for high-resolution medical images. To address these limitations, we propose TS-AC, an enhanced T–S framework for efficient abnormality detection in chest CT images. Our method introduces two novel modules: aggregation masked convolution (AMC) and concentrated linear attention (CLA). AMC encourages the model to reconstruct masked central regions using contextual cues during training, effectively mitigating overcorrection of anomalies. CLA incorporates depthwise separable convolution into linear attention, enhancing local feature modeling while significantly reducing computational cost compared with standard Softmax-based attention. Extensive experiments conducted on two representative chest CT datasets demonstrate that TS-AC outperforms state-of-the-art methods by 6.3% and 9.1% in AUC on the Zhanglab and CheXpert datasets, respectively. It also achieves consistent improvements in $F1$ score and accuracy. Furthermore, CLA contributes to a notable reduction in model complexity. Specifically, it reduces the number of parameters and FLOPs, while improving inference speed by more than 7% compared with conventional attention mechanisms. These results highlight the effectiveness and deployment potential of TS-AC in real-world clinical scenarios.