STD-YOLOv7:A small target detector for micronucleus based on YOLOv7

Abstract

The micronucleus of cells represents a form of abnormal structure in eukaryotic organisms. The detection of cellular micronuclei is applied in diverse aspects including the assessment of radiation-induced damage, experiments on new drugs, as well as the domain of food safety. Currently, however, research on micronucleus recognition remains limited, with detection accuracy often proving insufficient. In response to these challenges, we propose the STD-YOLOv7 micronucleus recognition algorithm, which integrates the YOLOv7 object detection framework with the Coordinate Attention (CA) mechanism and the Res-ACmix module, specifically tailored for recognizing cellular micronuclei. The CA mechanism enhances feature map expression, while the Res-ACmix module optimizes feature extraction. Both are applied within the feature extraction network, enabling refined feature transfer throughout the network. Furthermore, incorporating Dropout within the Backbone improves overall model performance by mitigating overfitting. Predictions are made at each layer’s prediction head to generate final results. Experimental results on the constructed SRCHD dataset show that the proposed STD-YOLOv7 algorithm surpasses other comparable methods in performance on this dataset and also performs well on publicly available datasets. On the SRCHD dataset, STD-YOLOv7 achieved significant improvements, including a 6.37 % increase in mean Average Precision (mAP@50), a 5.51 % boost in Recall, and a 5.01 % rise in Precision.