Segmentation of Medical Images Using Deep Learning and Texture Enhancement Based on Fractional Derivative Operators

This paper explores the task of automatic intracranial hemorrhage (ICH) segmentation based on computer tomography (CT) data, with potential applications in biomedical engineering. The main focus is, on improving segmentation accuracy by incorporating texture enhancement techniques based on fractional order derivatives. The study looks at the segmentation of ICH using $\mathrm{U}-\mathrm{Net}$, a deep learning model that is widely used in the field of the segmentation of medical images. The training process employs a parallel algorithm using CUDA technology. Afterwards, an investigation is conducted into a texture enhancement technique that relies on Riesz fractional order derivatives. The goal is to capture intricate details and subtle textures with the potential to enhance segmentation accuracy. To assess the impact of this preprocessing method on the automatic ICH segmentation problem, the U-Net model undergoes retraining and validation. The texture-enhanced images are analyzed to interpret the obtained results. The findings indicate a subtle yet discernible enhancement in accuracy, as gauged by the Jaccard and Dice coefficients. This emphasizes the auspicious potential of the explored texture enhancement method in improving intracranial hemorrhage segmentation within the realm of biomedical engineering.