An improved Yolov5s based on transformer backbone network for detection and classification of bronchoalveolar lavage cells

• Published in: Computational and structural biotechnology journal Vol. 21; pp. 2985 - 3001
• Main Authors: Wu, Puzhen, Weng, Han, Luo, Wenting, Zhan, Yi, Xiong, Lixia, Zhang, Hongyan, Yan, Hai
• Format: Journal Article
• Language: English
• Published: Netherlands Elsevier B.V 01.01.2023; Research Network of Computational and Structural Biotechnology; Elsevier
• Subjects: algorithms; biopsy; biotechnology; Bronchoalveolar lavage cells; Cell detection; Convolutional neural network; Deep learning; eosinophils; lungs; macrophages; neck; neural networks; neutrophils; Transformer; Deep learning; Bronchoalveolar lavage cells; Transformer; Convolutional neural network; Cell detection
• DOI: 10.1016/j.csbj.2023.05.008
• ISSN: 2001-0370

Biological tissue information of the lung, such as cells and proteins, can be obtained from bronchoalveolar lavage fluid (BALF), through which it can be used as a complement to lung biopsy pathology. BALF cells can be confused with each other due to the similarity of their characteristics and differences in the way sections are handled or viewed. This poses a great challenge for cell detection. In this paper, An Improved Yolov5s Based on Transformer Backbone Network for Detection and Classification of BALF Cells is proposed, focusing on the detection of four types of cells in BALF: macrophages, lymphocytes, neutrophils and eosinophils. The network is mainly based on the Yolov5s network and uses Swin Transformer V2 technology in the backbone network to improve cell detection accuracy by obtaining global information; the C3Ghost module (a variant of the Convolutional Neural Network architecture) is used in the neck network to reduce the number of parameters during feature channel fusion and to improve feature expression performance. In addition, embedding intersection over union Loss (EIoU_Loss) was used as a bounding box regression loss function to speed up the bounding box regression rate, resulting in higher accuracy of the algorithm. The experiments showed that our model could achieve mAP of 81.29% and Recall of 80.47%. Compared to the original Yolov5s, the mAP has improved by 3.3% and Recall by 3.67%. We also compared it with Yolov7 and the newly launched Yolov8s. mAP improved by 0.02% and 2.36% over Yolov7 and Yolov8s respectively, while the FPS of our model was higher than both of them, achieving a balance of efficiency and accuracy, further demonstrating the superiority of our model. [Display omitted]