FEW-VIEW X-RAY CT SEGMENTATION OF THIN TUBULAR STRUCTURES

Abstract

X-ray CT scanning is both time-consuming and radiation-intensive due to the large number of projections needed to produce high-quality images. Reducing the number of projections can accelerate acquisition and reduce radiation exposure, but often leads to degraded image quality, making the segmentation of fine-grained structures particularly challenging. To address this few-view imaging challenge, this study proposed a modified 3D U-Net model for segmenting thin tubular structures, with scalability that ensures high accuracy even with reduced input quality. The proposed model integrates residual blocks, squeeze-and-excitation (SE) blocks, and dilated convolutions to enlarge the receptive field and emphasise important channels, without significantly increasing the number of learnable parameters. Training was performed on few-view CT reconstructions generated from images with a dense projection. Experiments conducted in plant root and lung vessel datasets demonstrate that certain enhancements improved segmentation accuracy under severe sparsity.