Evaluation of Nano-Object Magnetization Using Artificial Intelligence

Despite the immense potential of magnetic nanoparticles in biomedicine, their widespread application is hindered by the lack of experimental methodologies for the widely available measure of fundamental physical parameters such as hydrodynamic radius and magnetization. In this study, we propose employing “artificial intelligence-based” image analysis to extract these parameters from videos of experiments with magnetic nanoparticle concentration under the influence of magnetic field. Various solutions of magnetic nanoparticles and their complexes with liposomes, each having different magnetization levels, display distinctive temporal dynamics. The videos captured during these experiments were used to create a dataset of time-dynamic video frames, which was then processed using machine-learning techniques. For semantic segmentation of analyzed frames within a specific experiment, neural network architectures based on U-Net have been trained. We achieve quite accurate predictions of nano-object relative magnetization (RM) within a percentage error range of 11.4% to 24.1% across a spectrum of nano-object magnetizations ranging from 12 to 68 emu/g.