Embedded System Application for Motion Prediction in Fluorescence Imaging of Live Cell Processes

Abstract

This thesis evaluates different algorithms for motion prediction in the context of live cell imaging using portable microscopy. The objective is to reduce both the time required for image acquisition and the resulting file size, by focusing only on areas containing relevant information. This approach improves image quality and data efficiency, making it suitable for low-power, embedded microscopy platforms where processing resources are limited. An often used solution is to integrate an X-Y motorised worktable that moves either the microscope or the sample in a preselected grid pattern. This allows for the system to capture a series of smaller images, which then can later be combined through image stitching to create a larger, high-resolution final image. However, in portable microscope systems, there are often limitations caused by the size, power consumption, and available processing capacity. These systems typically rely on embedded hardware with reduced computational performance. As a result, image acquisition becomes slower, which makes the observed motion of dynamic samples, such as living cells, appear faster. In these cases, it becomes important to predict and track the movement of the cells during scanning to ensure proper alignment and reduce data loss. When running on a desktop system, models like Decision Tree and Random Forest gave very low mean squared error (MSE) although Random Forest had a much longer training time. Simpler models like Ridge and Linear Regression were much faster to train and predict, but they had higher errors. Deep learning methods like LSTM and SVR had the worst performance in terms of time and resources. For example, SVR had an inference time of 134 seconds and the highest error values, showing that it's not suitable for low-resource systems. When these models were deployed on the Raspberry Pi 5 using TensorFlow Lite, the results also showed that lighter models were more efficient. For example, Linear Regression had a fast inference time and low CPU usage, while Random Forest and Gradient Boosting had better accuracy but required more CPU usage. More complex models like CatBoost and SVR had higher MSE and CPU usage, making them less ideal for real-time use on embedded systems. Overall, Random Forest gave a good balance between accuracy and processing load, making it one of the best options for motion prediction in portable microscope setups.