Influence of hyperparameters on the performance of deep learning-based microrobotic localization under phantom tissue

For the effective operation of medical microrobots within living organisms and precise targeting, it is imperative to employ imaging techniques closely integrated with real-time deep-tissue tracking methods. However, due to a typically low Signal-to-Noise ratio images with strong background, it is hard for traditional tracking methods to achieve sufficient accuracy. This challenge can be addressed by deep learning-based tracking with a real-time detection model. However, a multitude of design choices and Hyperparameters influence the performance. In this study we compared the influence of the hyperparameters and model architecture versions of the “you only look once” (YOLO) network. We use experimental data from a magnetic microrobot imaged with Photoacoustics through 5 mm phantom tissue to evaluate the tracking in comparison with an optical reference. The deep-learning based methods consistently achieved lower missing-detection ratios. Regarding the Root Mean Square localization error, we observed that increasing the weight of the box loss function and utilizing the distribution focal loss can enhance the performance by 10%. Furthermore, it can be seen that YOLOv9 consistently outperformed its predecessor YOLOv8. This study quantifies the robustness of deep-learning based tracking of medical microrobots under tissues.