MSQC-UNet: A UNet-based Multi-view Sputum Quality Classification Network

Tuberculosis (TB), one of the world’s deadliest infectious diseases, demands cost-effective screening solutions. While sputum smear microscopy is widely used, its accuracy heavily depends on sample quality. Current AI-based assessments relying on single-image analysis fail to capture spatial heterogeneity, leading to inconsistent diagnoses. To address this, we propose MSQC-UNet, a novel framework integrating U-Net’s segmentation capability with a multi-view attention fusion mechanism to comprehensively analyze smear spatial features. Experiments on clinical datasets show MSQCUNet achieves 97.8% accuracy (2% higher than baseline), with robustness to staining variations. This study provides a reliable, automated tool for TB screening, particularly in resource-limited settings.