Ultrasensitive detection of low-abundance protein biomarkers is essential for early disease diagnosis and therapeutic monitoring. While droplet digital enzyme-linked immunosorbent assay (ddELISA) addresses this need by enabling attomolar sensitivity, its performance remains limited by conventional image analysis methods, restricting accurate high-throughput image analysis. Herein, we developed a custom deep learning model, YOLO-Drop, for accurate, high-throughput analysis of droplet images and deployed it on an NVIDIA Jetson Orin Nano embedded platform equipped with an intuitive graphical user interface (GUI) to support real-time and user-friendly operation, thereby significantly enhancing the speed, automation, and accuracy of ddELISA. Built upon the YOLOv8 architecture, the YOLO-Drop model was customized with deformable convolution (DConvModule and DC2f) and BiFormer modules, a high-resolution feature pyramid network (HR-FPN), a small-object prior (SOP), and a class-aware nonmaximum suppression (CA-NMS) to enhance small-object recognition in complex ddELISA droplet images. Further trained with a high-quality annotated data set of 5,574 droplet images containing ∼750,000 droplets with diverse signal patterns and intensities, the YOLO-Drop model achieved high detection accuracy (99.69%) across heterogeneous ddELISA droplet images. Such high detection accuracy, together with fast inference on the Jetson platform, enables YOLO-Drop to perform reliable, real-time, on-device droplet image analysis. When applied to ddELISA, YOLO-Drop enabled the assay to detect interleukin-6 (IL-6, as a representative biomarker) down to 9.57 aM with excellent performance in complex biological matrices. This work underscores the transformative potential of deep-learning-assisted data analysis in advancing next-generation biosensing platforms toward accurate, automated, and high-throughput biomarker quantification in clinically relevant settings.
Xu et al. (Sun,) studied this question.