Efficiency of Kolmogorov-Arnold Networks in Small Medical Samples (Case Study of 2D Brain MRI Image Segmentation)

The aim of the study was to evaluate the efficiency of the KANU-Net 2D architecture based on U-Net in medical segmentation tasks of 2D brain MRI images on BraTS dataset with a limited number of training samples.

The experiments were carried out using the subsamples containing 50, 100, and 150 images. The study described the data preprocessing steps, including normalization, gamma correction, cropping, and augmentation. A combination of Dice loss and BCE loss was used as a loss function. The network was optimized using AdamW. The network operation performance was evaluated using Accuracy and Dice coefficient for each region and its mean Dice.

KANU-Net 2D was experimentally demonstrated to achieve competitive performance comparable to current SOTA models of convolutional neural networks when trained on small samples. Specifically, the mean Dice coefficient reached 0.851 when using 100 training samples.

The conducted studies showed KANU-Net 2D network to outperform the Med-DANet segmentation model both in terms of a mean value and region classes. The model effectiveness for different tumor regions highlighted the ability of the KAN-based (Kolmogorov-Arnold network) approach to adapt to various image characteristics in medical segmentation tasks. The obtained results demonstrated the undeniable promise of applying KAN for medical image segmentation in small samples and can lay the foundation for further research in this field.