The Effect of Preprocessing on U-Net for Bladder Segmentation in CT Images

Abstract

This research proposes preprocessing techniques for computed tomography (CT) slices with the aim of improving the performance of a deep-learning segmentation model (U-Net model). The preprocessing techniques used in this study include window leveling, histogram equalization, Gaussian blurring, and cropping (incorporating mathematical morphology and histogram projection). The U-Net model is applied to three groups of input data sets (B-I to B-III) for training, validation, and testing. The segmentation performance is evaluated using metrics such as Dice similarity coefficient (DSC) and intersection over union (IoU). The trained U-Net model achieves the highest DSC (95.15%) and IoU (91.09%) under B-III dataset, utilizing cropped and enhanced CT input datasets. Window leveling, histogram equalization, Gaussian blurring, and cropping are identified as the optimal preprocessing techniques for bladder segmentation. This novel research applies diverse image processing techniques in medical image preprocessing to enhance the deep-learning U-Net model's segmentation performance.