سامانه همایش‌ها TSTA | ثبت‌نام و ارسال مقاله کنفرانس‌ها و کنگره‌های ملی

Home / سی و دومین کنفرانس ملی و دهمین کنفرانس بین المللی مهندسی زیست پزشکی ایران

Lightweight 3D U-Net for Robust Liver Segmentation in Multi-Institutional CT Datasets

Authors :

Seyyed Mohammad Hosseini¹ Faeze Salahshour² Ahmadreza Sebzari³ Masoomeh Safaei⁴ Hossein Ghadiri Harvani⁵

1- دانشگاه علوم پزشکی تهران 2- دانشگاه علوم پزشکی تهران 3- دانشگاه علوم پزشکی بیرجند 4- دانشگاه علوم پزشکی تهران 5- دانشگاه علوم پزشکی تهران

Keywords :

Liver،Segmentation،Computed Tomography (CT)،3D U-Net

Abstract :

A computed tomography (CT) image of the liver and surrounding structures provides detailed cross-sectional images, which highlight anatomical variations and pathological conditions. The combination of CT and U-Net networks is a well-known method for liver segmentation, which is vital for accurate diagnosis, treatment planning, and surgical intervention. However, the high computational demands of recent 3D U-Net–based architectures prevent their deployment in resource-constrained environments. A lightweight 3D U-Net optimized for liver segmentation is proposed in this study, maintaining high performance while reducing computational complexity drastically. Several institutional datasets of 250 abdominal CT volumes were compiled from public benchmarks (LiTS, IRCAD) and local clinical sources, encompassing anatomical, pathological, and protocol variations. An isotropic resampling procedure was used to resample, normalize intensity, standardize crops, and augment data on-the-fly. With fewer than two million parameters, the proposed model retains the encoder-decoder and skip-connection designs of conventional 3D U-Nets. An evaluation of a 30% independent set of tests achieved Dice similarity coefficients of 0.85 ± 0.02, intersect-over-unions of 0.82 ± 0.03, inference times under 0.7 s and GPU memory consumption below 2 GB. The performance was consistent across public and local datasets, highlighting the importance of heterogeneous training data. Even though the proposed model is slightly less accurate than heavy architecture, it delivers near-real-time segmentation with minimal resource consumption, so it can be integrated into clinical workflows, especially in environments where computational resources are limited.