Cell Growth Assessment Using Segmentation Maps for Clone Screening

The patent replaces the manual mechanical visual inspection process with an automated image analysis system using machine learning algorithms. The system captures images of cell cultures and uses trained neural networks to automatically detect, segment, and count cells, substituting human analysts with computational algorithms that process images rapidly and consistently.

The patent introduces an intermediary layer of automated image processing and machine learning models between the raw well images and the final growth assessment. This intermediary system includes preprocessing steps, cell detection algorithms, segmentation networks, and counting mechanisms that translate visual data into quantitative growth metrics without direct human intervention.

The patent employs segmentation techniques to divide the image processing task into distinct stages: initial cell detection, individual cell segmentation from the background and from each other, and final counting. By breaking down the complex task of counting dense colonies into manageable segmentation steps, the system maintains accuracy even when cells are tightly packed together.

The patent applies partial action by focusing the automated system on specific measurable outcomes (colony presence, relative growth trends) rather than attempting perfect individual cell counting in all scenarios. For dense colonies where precise counting is extremely difficult, the system uses alternative metrics such as colony area measurement or presence/absence detection, which remain accurate even when individual cell boundaries are indistinct.

The patent implements self-service mechanisms where the system automatically performs tasks that would otherwise require manual intervention. The automated annotation feature allows the system to generate training data by having users review and correct only the most uncertain predictions, rather than requiring manual annotation of every training image. This self-correcting approach significantly reduces the labor intensity of creating training libraries.

The patent incorporates feedback loops where the automated analysis system continuously learns from user corrections and validation. User interactions with the system provide feedback that refines the machine learning models over time, improving accuracy while reducing the initial burden of training library creation. The system adapts to specific laboratory conditions and cell types based on this feedback.