Enhancing Medical Image Segmentation using Hexagonal Convolutional Neural Networks

This paper introduces Hex U-Net, an innovative segmentation framework
designed for hexagonal pixel-based images. It overcomes the shortcomings
of conventional methods that rely on rectangular pixel-based representations, by
leveraging a hexagonal tessellation structure to better capture the complexities
inherent in medical images. By harnessing recent advancements in pre-processing
techniques and convolutional neural network architectures, the study systematically
explores the integration of hexagonal grid structures into deep learning
frameworks, with the objective of enhancing algorithmic performance and
dice score in medical image segmentation tasks. The implementation process involves
a series of steps, including dataset pre-processing to convert the image
data into hexagonal pixel-based format, the introduction of hexagonal convolution
kernel, and development of the Hex U-Net architecture. Experimental results
demonstrate significant improvements in segmentation accuracy, with Hex UNet
achieving an impressive dice score of 0.873 compared to conventional architectures
such as U-Net, VGG16, and ResNet18. The superior performance of Hex
U-Net underscores the efficacy of hexagonal pixel-based representations and specialised
convolutional operations in capturing intricate spatial relationships
within medical images, thereby offering promising prospects for precise diagnosis
and treatment planning in clinical practice.