Micro-object Detection in Microfluidic Devices Using Machine Learning

The patent replaces traditional mechanical/image processing detection methods with machine learning algorithms (convolutional neural networks). The system uses trained neural networks to automatically identify micro-objects by learning their characteristics from training data, substituting complex image processing pipelines with a trained model that achieves higher accuracy despite the translucent nature of micro-objects and non-uniform backgrounds.

The patent transforms the detection approach by changing from direct image analysis to using trained machine learning models with multiple parameters. The system adjusts detection parameters dynamically based on training data, allowing the model to adapt to varying conditions such as different micro-object types, backgrounds, and imaging conditions, thereby improving detection accuracy.

The patent performs preliminary training of machine learning models offline using extensive training datasets. Once trained, the models are saved and can be deployed for rapid inference. This preliminary action separates the time-consuming training phase from the actual detection phase, allowing fast processing during runtime while maintaining high accuracy through pre-learned features.

The system uses partial action by focusing the machine learning model only on the most relevant features for detection rather than processing all image data equally. The convolutional neural network automatically learns to attend to critical regions and features, reducing unnecessary computational overhead while maintaining detection accuracy.

The patent applies local quality by training the machine learning model to recognize specific local features and patterns that distinguish micro-objects from background elements. The convolutional neural network learns to focus on local characteristics such as edge patterns, texture variations, and spatial relationships that are unique to micro-objects, enabling reliable detection even when background features are similar in size.

The trained machine learning model acts as an intermediary between the raw image data and the detection output. It processes the complex relationship between micro-objects and non-uniform backgrounds by learning intermediate representations that highlight distinguishing features, thereby reducing false detections and missed detections through learned feature discrimination.