CADD-UNet : 3D Coordinate Attention with Double Dense Connection for Ureter Segmentation

Abstract

Over the past decade, computed tomographic urography has emerged as the primary imaging modality for diagnosis the ureter in clinical environment. In the human body, Ureter appears to be a small structure with a cross section of 3~4mm in CT axial scan, both its shape and intensity are similar to blood vessels in non-contrast CT. Therefore, contrast agents are usually injected to accurately diagnose ureter. However, patients with reduced renal function or severe side effects of contrast agents take CT scan without injection makes hard to diagnosis ureter. In this paper, research conducted to find whole ureter path in non- contrast CT by using deep learning. We introduce a novel loss function, called Portion Loss, which prevents small object from being ignored when training with larger sizes. Additionally, we introduce CADD-UNet, which continuously provide the feature before convolution operation to give information about small object that may have been lost. And training different receptive fields through various dilated patterns. Portion Loss improves the performance of ureter segmentation by 4.3% over DiceCE loss. CADD-UNet improves 1.0% over UNet. Lastly, we point out that dice scores are mainly used for evaluating medical image segmentation research, but that evaluation metrics do not necessarily have to be pixel-wise accurate when clinically judged, and propose “Distance score” that clinician can judge more intuitively.