Efficient Multi-Label Attribute Classification and Recognition of Vaginitis Bacteria Based on Deep Learning and model fine-tuning

Abstract

Bacterial vaginosis (BV) is the most common gynecological complaint affecting the health of a large percentage of women worldwide. Traditional manual microscopy methods are expensive and time-consuming, to improve accuracy and efficiency, automated bacterial identification devices with detection intelligence algorithms are urgently needed. We propose a Fine-tuned SmallerVGG (FTS-VGG) convolutional network model-based multi-label classification method for bacteria. Comparison experiments were deployed on several basic backbone networks, including transfer learning on pre-trained VGG19, demonstrating that the proposed method achieves the advantages of being lighter, faster, more accurate, and more efficient. Due to the high cost of time and expertise of experienced clinicians, we use random erasing for data augmentation to address the challenge of dataset collection and annotation, experiments demonstrate its robustness to occlusion. The proposed method has theoretical and practical implications and the potential to be widely extended to other microscopic imaging applications.