2026 Volume 15 Pages 398-407
With advances in computer vision, medical image segmentation has become a crucial tool for diagnosis and treatment of diseases. U-Net, a widely used deep learning model, excels in extracting anatomical structures from medical images; however, optimization for multiple anatomical targets within a single framework can be challenging, particularly in cadaveric imagery with heterogeneous structures. To address this, we propose a parallel U-Net that employs multiple U-Net models to simultaneously extract various anatomical structures―arteries (abdominal aorta), veins (inferior vena cava), and organs (kidneys)―from static and dynamic (video) images. Each U-Net model is trained separately on labeled anatomical data, and their outputs are merged to generate multistructure segmentations. Experiments were conducted using abdominal images obtained from 11 Japanese cadavers. In this system, the four-layer U-Net optimized the balance between extraction accuracy and time. It achieved pixel-wise accuracy values exceeding 0.95 in most cases, even on images of cadavers not used in training. Additionally, real-time segmentation was successfully performed at over 10 fps on video frames, enabling simultaneous estimation of multiple structures despite variability across cadavers. In conclusion, the proposed parallel U-Net effectively segments multiple abdominal structures in both still and video images, offering a promising approach to advance medical imaging.
