Active-matrix electrowetting-on-dielectric (AM-EWOD) system, integrated with hundreds of thousands of active electrodes, can simultaneously realize multiple on-chip biochemical reactions at the single-cell level. An intelligent detection system is critical for fully automating manipulations of thousands of digitalized bio-samples and programming the subsequent experiments in real time. Conventional image processing algorithms are sensitive to factors such as lighting and noise, resulting in limited generalizability. They lack the capability to autonomously learn features and often require manual parameter adjustments. In this work, we developed a series of deep learning algorithms based on an AM-EWOD system for sample detections. We used the U-net model to quantitatively evaluate different splitting methods on sample droplet generation uniformity. The results revealed that droplets generated using the “one-to-two” strategy exhibits optimal uniformity. We used the YOLOv5 model to monitor the droplet splitting success rates over 18 different AM-EWOD chips, and a 97.7% splitting success rate was observed. The results indicated that the model precision was 99.980% and the model recall was 99.976% through manual verification. In addition, we used an improved YOLOv8 model to detect single cells in nanoliter droplets effectively. In comparison with manual verification, the results showed that the model achieved a precision of 99.260% and a recall of 99.193%. By leveraging an artificial intelligence enabled smart detection system, AM-EWOD system has shown great potential as a ubiquitous platform for true lab-on-a-chip.