Adaptive region-oriented masked vision retentive network for predicting macrovascular invasion in hepatocellular carcinoma

The aim of the present study was to develop the Adaptive Region-Oriented Masked Vision Retentive Network (AROMA ViR) model, which can efficiently learn the morphological structures of the liver, to predict macrovascular invasion (MI) in hepatocellular carcinoma (HCC). Retrospective CT images were obtained from the University of Texas MD Anderson Cancer Center, in accordance with The Cancer Imaging Archive data usage policy and restrictions. The image dataset comprised 51,968 slices taken during the arterial phase in 105 patients with HCC. We split the data patient-wise into training, validation, and test datasets in a 6:2:2 ratio after applying specific exclusion criteria. The AROMA ViR was designed to enhance the relevant areas in the retention map by incorporating spatial information for liver parenchyma, tumor, and portal vein. The model applied causal masks specialized for specific liver shapes for each slice image into retention encoders. We compared the proposed model with Residual Network 101, Vision Transformer, and Vision Retentive Network. We calculated the area under the receiver-operating characteristic curve (AUC-ROC) and that under the precision recall curve (AUC-PR). We also obtained accuracy, sensitivity, specificity, and F1-score using Youden’s index. AROMA ViR pretrained by ImageNet showed AUC-ROC of 0.860, AUC-PR of 0.790, accuracy of 0.853, sensitivity of 0.719, specificity of 0.902, and F1 score of 0.578.