Deep learning-based automated segmentation of liver and tumors on CT arterial portography and hepatic arteriography for Y-90 radioembolization

Y-90 radioembolization (RE) is an effective treatment for inoperable liver tumors. Pre-treatment planning using Tc-99m MAA requires segmentation of normal liver (NL), lungs and tumors. Deep learning (DL)-based automatic segmentations have been developed for liver tumor segmentation usually based on the portal phase contrast CT. However, CT arterial portography (CTAP) and CT hepatic arteriography (CTHA) are essential for RE to provide more structural and blood supply information of tumors. This study aims to develop a DL-based method for automatic NL and tumor segmentation using both CTAP and CTHA images for Y-90 RE treatment planning. Tc-99m-MAA SPECT/CT, CTAP and CTHA images of 48 patients were retrospectively collected. NL and tumor maps were provided as the gold standard. A self-adapting framework, nnU-Net, with 6-fold cross-validation was performed to evaluate 4 training strategies: (1) Att-nnUNet
CTAP+CTHA
: nnU-Net with attention-gate trained on CTAP and CTHA; (2) nnUNet
CTAP+CTHA
: nnU-Net trained on CTAP and CTHA; (3) Att-nnUNet
CTAP
: nnU-Net with attention-gate trained on CTAP; and (4) nnUNet
CTAP
: nnUNet trained on CTAP. The segmentation performance was evaluated using Dice. The DL-segmented tumor and NL maps were applied to the corresponding SPECT for calculating the tumor-to-NL ratio (TNR), and absorbed dose of NL, D
NL
, and tumor, D
T
. All training strategies achieved a mean Dice >0.93 for NL. Att-nnUNet
CTAP+CTHA
had the best Dice with 0.83±0.12 followed by nnUNet
CTAP+CTHA
with 0.82±0.11, while network trained by only CTAP achieved mean Dice <0.73 for tumor. Att-nnUNet
CTAP+CTHA
had the least mean absolute error of TNR (13.03%±17.73%), D
NL
(7.09%±11.50%) and D
T
(8.41%±12.85%). DL-training on both CTAP and CTHA can achieve better segmentation performance compared with CTAP alone. The attention module can further improve segmentation accuracy.