DEVELOPMENT OF BESSEL POLYNOMIAL-BASED KOLMOGOROV-ARNOLD NETWORKS WITH APPLICATION TO MAPPING MAGNETIC RESONANCE SIGNAL FROM HUMAN BRAIN METABOLITES

ABSTRACT This project proposes the BesselKAN model, a polynomial-based Kolmogorov Arnold Network architecture designed for accurate MRI image analysis. The model combines the strengths and ability of Kolmogorov-Arnold-Arnold Networks (KAN) and Bessel polynomials to accurately approximate the target function (Bloch equation), an important constituent in MRI imaging. The traditional MLP often fall short in approximating complex equations or signals. This work presents the BesselKAN model that integrate polynomial degree as layers to simulate signals. The model was trained and evaluated using the gray matter and white matter in vivo dataset. Key metrics, the training loss was used to assess and compare the model performance. Over the last epoch, the BesselKAN achieved the score: Training Loss: 0.0327 in comparison, the MLP model recorded Training Loss 0.0359. The results showed that the BesselKAN has the ability to perform and even outperform the traditional Multilayer Perceptron (MLP) model. Using the 2D convergence plot, contour plots and 3D plots the BesseKAN demonstrate greater and good accuracy, improved generalization, and enhanced robustness. This project contributes to the field of medical imaging analysis by introducing the BesselKAN model as a powerful tool for MRI image analysis, validating the model's performance on a comparative analysis with the MLP model. This research and work opens a way for future innovations within medical imaging and underlines the capability and potency of the BesselKAN.