A semi-supervised TB detection model based on improved YOLOv5

Existing object detection methods often fail to detect Tubercle Bacill (TB) in pathological images because of its small size and complex background. To improve the detection accuracy, we develop an improved target detection algorithm based on YOLOv5s, called YOLOv5-TB. More concretely, to deal with the small scale and limited number of TB, a pasting strategy is used to enhance the ability to learn features related to small targets. In addition, we use a self-training semi-supervised learning strategy to train the model, aiming to take advantage of a large amount of unlabelled data to generate the required labeled data. Besides, we add a detection head to the model specifically for detecting small targets to overcome the limitations related to the prediction frame size, which affects multi-scale feature extraction when the target scale varies significantly. Finally, to validate the effectiveness of our proposed model, we create a dataset of 2550 images of TB. The experimental results show that the proposed method achieves an impressive 98.5% accuracy in TB detection, which demonstrates the superiority of our method.