Machine learning approach for enumeration of circulating cells with diffuse in vivo flow cytometry.

SIGNIFICANCE: Diffuse in vivo flow cytometry (DiFC) is an emerging technique for enumerating rare, fluorescently labeled circulating tumor cells (CTCs) in small animals without drawing blood samples. DiFC uses detection of transient fluorescent peaks in time-series data. Previously, we used a simple amplitude threshold-based method for identifying peak candidates, but this ignored potentially useful information in peak shape that could reduce false-positive detections and increase detection efficiency of lower-amplitude peaks. AIM: Our aim is to develop a machine learning (ML)-integrated signal processing approach for DiFC that improves CTC enumeration by better discriminating CTC peaks from artifacts. APPROACH: We developed an ML-integrated approach that incorporates a convolutional neural network (CNN) classifier. The CNN was trained to distinguish CTC peaks from artifacts by analyzing peak amplitude and temporal shape characteristics. Performance was evaluated on in silico, control, and CTC-bearing mouse datasets. RESULTS: The CNN classifier achieved accuracy, precision, sensitivity, and specificity exceeding 94% on test data. Compared with our previously published threshold-based approach, the ML-integrated method increased the number of correctly identified CTCs and their flow direction while reducing false detections across evaluation datasets. CONCLUSIONS: The ML-integrated approach substantially improves DiFC CTC enumeration, enabling robustness against artifacts in noisy conditions.