Intelligent larval zebrafish phenotype recognition via attention mechanism for high-throughput screening

• Published in: Computers in biology and medicine Vol. 188; p. 109892
• Main Authors: Wang, Baihua, Sun, Qi, Liu, Yujia, Zhang, Jiheng, Li, Gaozheng, Wu, Sifang, Zheng, Houbing, Ye, Jialin, Zhou, Meihua, Zheng, Haisu, Yu, Yongqiang, Zhong, Yi, Wu, Yuanzi, Huang, Da, Wang, Biao, Weng, Zuquan
• Format: Journal Article
• Language: English
• Published: United States Elsevier Ltd 01.04.2025; Elsevier Limited
• Subjects: Abnormalities; Accuracy; Animals; Attention; Attention mechanism; Automation; Classification; Danio rerio; Data augmentation; Data collection; Data processing; Datasets; Deep Learning; Defects; Edema; Genotype & phenotype; High-throughput screening; High-Throughput Screening Assays - methods; Image Processing, Computer-Assisted - methods; Internal Medicine; Labels; Laboratories; Larva; Larval zebrafish; Microscopy; Morphology; Neural networks; Other; Pharmacovigilance; Phenotype; Phenotype classification; Phenotypes; Potassium; Product safety; RECNet; Toxicology; Zebrafish; High-throughput screening; Attention mechanism; Larval zebrafish; RECNet; Phenotype classification
• ISSN: 0010-4825; 1879-0534; 1879-0534
• DOI: 10.1016/j.compbiomed.2025.109892

Larval zebrafish phenotypes serve as critical research indicators in fields such as ecotoxicology and safety assessment since phenotypic defects are closely related to alterations of underlying pathway. However, identifying these defects is time-consuming and requires specialized knowledge. We proposed a deep network model called RECNet, which combines attention mechanisms and residual structures. In terms of data processing, we applied the mixup data augmentation technique and accumulated a collection of 6805 larval zebrafish phenotype images, mostly generated from our laboratory. Our proposed model was deployed to execute two distinct tasks, including a four-classification of zebrafish phenotypes and a seven-classification involving mixed labels for abnormalities. In the four-class classification task, the RECNet model achieved an accuracy of 0.949, with a mean area under the curve of 0.986 and an F1-score of 0.966. Through interpretable research, attention mechanisms enable the model to focus more accurately on regions of interest. In the mixed-label seven-classification task for anomalies, our model achieved an accuracy of 0.913 and a mean average precision value of 0.847 by employing the weighted loss function (DFBLoss). Furthermore, in a new test dataset, the RECNet model achieved accuracy rates of 0.924 and 0.876 for the two tasks, respectively. Our RECNet model was trained by orders of magnitude larger dataset than previous studies and also showed better accuracy rates. Our method holds promise for diverse applications within zebrafish laboratories and fields such as toxicology, providing indispensable support to scientific research. •A total of 6805 zebrafish phenotype images were collected from the laboratory and relevant literature.•The RECNet model and the DFBLoss function were proposed for zebrafish phenotype recognition and classification.•The explanation tools demonstrated that the proposed model better focuses on the target regions of zebrafish phenotypes.•The proposed model outperforms other mainstream models and shows excellent recognition results on new test datasets.