Identification of macrotrabecular-massive hepatocellular carcinoma through multiphasic CT-based representation learning method.

Background: Macrotrabecular-massive hepatocellular carcinoma (MTM-HCC) represents an aggressive subtype of HCC and is associated with poor survival.

Purpose: To investigate the performance of a representation learning-based feature fusion strategy that employs a multiphase contrast-enhanced CT (mpCECT)-based latent feature fusion (MCLFF) model for MTM-HCC identification.

Methods: A total of 206 patients (54 MTM HCC, 152 non-MTM HCC) who underwent preoperative mpCECT with surgically confirmed HCC between July 2017 and December 2022 were retrospectively included from two medical centers. Multiphasic radiomics features were extracted from manually delineated volume of interest (VOI) of all lesions on each mpCECT phase. Representation learning based MCLFF model was built to fuse multiphasic features for MTM HCC prediction, and compared with competing models using other fusion methods. Conventional imaging features and clinical factors were also evaluated and analyzed. Prediction performance was validated by ROC analysis and statistical comparisons on an internal validation and an external testing dataset.

Results: Fusion of radiomics features from the arterial phase (AP) and portal venous phase (PAP) using MCLFF demonstrated superior performance in MTM HCC prediction, with a higher AUC of 0.857 compared with all competing models in the internal validation set. Integration of multiple radiological or clinical features further improved the overall performance, with the highest AUCs of 0.857 and 0.836 respectively achieved in the internal validation and external testing set.

Conclusions: Multiphasic radiomics features of AP and PVP fused by the MCLFF have demonstrated substantial potential in the accurate prediction of MTM HCC. Clinical factors and Radiological features in mpCECT contribute incremental values to the developed MCLFF strategy.