AD-10: Machine Learning Approaches for Magnetic Nanoparticles Applications in Biomedicine

Machine learning (ML) is emerging as a valuable tool for building models of large datasets of highly interdependent, highly non-linearly coupled data. Magnetic systems, because of their intrinsic non linearity and complex interplay among many different quantities, are particularly suitable for being analysed with such approaches, provided large and coherent datasets are available to feed the machine learning model. In this work, we present two examples of machine learning approaches applied to magnetic nanoparticles of interest in biomedicine. With the first approach [1], we exploit datasets of magnetic nanoparticles (MNP) properties built through numerical simulations. The simulated MNPs and the environment in which they are supposed to operate can represent different scenarios of interest for biomedical applications, such as magnetic hyperthermia or heat-induced drug release. Input quantities, such as MNPs diameter, magnetic anisotropy, magnetic saturation, mutual interactions, space distribution, or operating conditions such as temperature, applied field maximum intensity (vertex value) and frequency are used to numerically calculate the MNPs response through a two-wells numerical model, and output values as coercivity, remanence, loop area are extracted and used to build the dataset. Finally, a ML model exploits the simulated dataset to predict the behaviour of the MNPs in conditions or for values of their properties that were not originally present in the dataset, but that may turn out to be valuable in biomedical applications. The second approach addresses the Magnetic Particles Imaging (MPI) technique, where magnetic nanoparticles with typical diameters in the 10-20 nm range are driven in correspondence of tumour masses and are triggered with a static magnetic field gradient to form a sensitive region that is excited to produce a measurable signal by an external, low-intensity ac electromagnetic field. By scanning with a suitable device (an antenna) over the patient’s body, maps of the magnetic response of the magnetic nanoparticles can be obtained, which can be used to identify the cancer masses and characterise some of their properties. With respect to other imaging diagnostic techniques, MPI does not use ionising radiation or intense magnetic fields, and is therefore attracting much attention. The solution of the direct problem, i.e. the calculation within the sensitive region of the real and imaginary parts maps given the initial nanoparticles distribution and their magnetisation vs. field curve, is time consuming but can be solved numerically to build a large dataset, whereas the inverse problem, i.e. the calculation of the nanoparticles distribution from the real and imaginary parts maps, is not trivial and may be severely affected by incomplete or inaccurate knowledge of the physical properties of the nanoparticles. Nonetheless, it is the inverse problem that is mostly relevant for diagnostic applications. We approach its solution with a machine learning model based on convolutional neural networks that reconstructs the magnetic nanoparticles distribution from their simulated third-harmonics susceptibility data. Both examples provide an insight on the power and versatility of ML approaches in investigating complex, highly nonlinear problems typical of magnetic systems, such as those involving magnetic nanoparticles, in the field of biomedical applications.References: [1] M. Coïsson, G. Barrera, F. Celegato, P. Allia, P. Tiberto, "Specific loss power of magnetic nanoparticles: a machine learning approach", Appl. Phys. Lett. Materials 10 (2022) 081108, doi 10.1063/5.0099498