A Lightweight and Robust Framework for Circulating Genetically Abnormal Cells (CACs) Identification Using 4-Color Fluorescence In Situ Hybridization (FISH) Image and Deep Refined Learning.

Circulating genetically abnormal cells (CACs) constitute an important biomarker for cancer diagnosis and prognosis. This biomarker offers high safety, low cost, and high repeatability, which can serve as a key reference in clinical diagnosis. These cells are identified by counting fluorescence signals using 4-color fluorescence in situ hybridization (FISH) technology, which has a high level of stability, sensitivity, and specificity. However, there are some challenges in CACs identification, due to the difference in the morphology and intensity of staining signals. In this concern, we developed a deep learning network (FISH-Net) based on 4-color FISH image for CACs identification. Firstly, a lightweight object detection network based on the statistical information of signal size was designed to improve the clinical detection rate. Secondly, the rotated Gaussian heatmap with a covariance matrix was defined to standardize the staining signals with different morphologies. Then, the heatmap refinement model was proposed to solve the fluorescent noise interference of 4-color FISH image. Finally, an online repetitive training strategy was used to improve the model's feature extraction ability for hard samples (i.e., fracture signal, weak signal, and adjacent signals). The results showed that the precision was superior to 96%, and the sensitivity was higher than 98%, for fluorescent signal detection. Additionally, validation was performed using the clinical samples of 853 patients from 10 centers. The sensitivity was 97.18% (CI 96.72-97.64%) for CACs identification. The number of parameters of FISH-Net was 2.24 M, compared to 36.9 M for the popularly used lightweight network (YOLO-V7s). The detection speed was about 800 times greater than that of a pathologist. In summary, the proposed network was lightweight and robust for CACs identification. It could greatly increase the review accuracy, enhance the efficiency of reviewers, and reduce the review turnaround time during CACs identification.