A Lightweight Deep Learning Framework for Identifying Circulating Genetically Abnormal Cells in Four-color Fluorescence in Situ Hybridization Images

Circulating abnormal cells are important biomarkers for cancer diagnosis and prognosis. It is characterized by safety, low cost and high reproducibility, and can be used as an important reference for clinical diagnosis. However, the identification and counting of circulating abnormal cells require manual operation, which is error-prone and time-consuming. In this study, a lightweight deep learning model was developed to automatically detect circulating abnormal cells in patients’ peripheral blood based on the features of clinical data. In addition, a difficult sample training strategy based on repetitive learning is proposed to help the model better distinguish between broken, weak, and adjacent signals in images. The study collected 44,000 four-channel fluorescent staining core data for training and testing the proposed model. The results on the testing set show that the model can identify four staining channels with accuracy higher than $95 \%$, precision, recall and $F 1$ higher than $96 \%$, AP50 higher than $97 \%$, and the number of parameters is significantly less than other models. In conclusion, the proposed model can quickly and sensitively identify circulating abnormal cells, which is clinically valuable for determining prognosis and can be used to diagnose metastasis.