ConvNeXt-2U: A 3-D Deep Learning-Based Segmentation Model for Unified and Automatic Segmentation of Lungs, Normal Liver and Tumors in Y-90 Radioembolization Dosimetry

Abstract

Y-90 radioembolization (RE) is an effective treatment for inoperable liver tumors. Pretreatment planning using Tc-99m-macroaggregated albumin (MAA) SPECT/CT requires segmentations of lung, normal liver and tumor, which could be delineated on low dose CT (LDCT), CT arterial portography (CTAP) and CT hepatic arteriography (CTHA). This study aims to develop a deep learning-based method for automatic lung, normal liver, and tumor segmentation for Y-90 RE treatment planning. Sixty-four sets of Tc-99m-MAA SPECT/CT, CTAP and CTHA images were retrospectively collected. Ground truth maps were provided by an experienced radiologist. We proposed ConvNeXt-2U, utilizing two U-Nets with connected skip connections and 3-D ConvNeXt blocks for joint segmentations. The LDCT, CTAP and CTHA were input to the two U-Nets. U-Net, attention U-Net, ResU-Net, MedNeXt, UNETR and Swin-UNETR were implemented for comparison. The segmentation performance was evaluated using Dice, Hausdorff distance (HD)95% and volume similarity (VS), and Y-90 RE dosimetrics, i.e., tumor-to-normal-liver ratio, lung-shunt fraction (LSF), absorbed dose (AD) of lungs, normal liver and tumors, and injected activity (IA). ConvNeXt-2U achieved the best performance in all segmentation indices and dosimetrics, except for HD95% of normal liver. It achieved mean Dice of 0.99, 0.93 and 0.77 in lungs, normal liver and tumors. ConvNeXt-2U provides a one-stop platform for unified segmentations for Y-90 RE treatment planning.