Towards multi-view sputum smear quality classification

Tuberculosis is a severe infectious disease affecting millions of people worldwide. Sputum smear microscopy, as a preliminary screening method for TB, is favored for its low cost and simplicity of operation. The quality of sputum smears directly influences the outcomes of microscopy. Moreover, the manual screening process carries a risk of infection. Currently, the automatic quality classification of sputum smears using computer vision and artificial intelligence technologies can significantly reduce manual labor. However, existing methods rely solely on single images, which are insufficient for capturing 3D information of specimens, thereby limiting the accuracy. In this paper, we propose GVUnet, a multi-view approach for evaluating sputum smear quality by integrating hierarchical feature extraction with 3D feature fusion. The network extracts shallow features to identify boundaries and contours, and deep features to capture depth and morphological features. Utilizing a Deep-Group-Shape architecture, GVUnet effectively fuses the deep features to generate a 3D feature. To validate the effectiveness of GVUnet, we designed a multi-view imaging instrument and collected artificial and natural datasets. The comparative experiments show that GVUnet achieved 98.5% accuracy on the artificial dataset. For the natural dataset, partial fine-tuning achieved 98% accuracy, while full fine-tuning reached 96% accuracy with a small sample size. Ablation studies further demonstrated that multi-view data are the key factor for improving accuracy. In conclusion, we adopted a multi-view method for sputum smear quality classification, which provides a 3D feature for analysis. This approach not only improves model performance but also offers a solution for practical applications.