2024 ARRS ANNUAL MEETING - ABSTRACTS

RETURN TO ABSTRACT LISTING


E4706. Prediction of Microvascular Invasion in HCC Based on Multimodality MR Radiomics and Delta Radiomics Model
Authors
  1. Yifan Pan; The School of Medical Imaging, Fujian Medical University
  2. Shunli Wang; The School of Medical Imaging, Fujian Medical University
  3. Yamei Liu; The School of Medical Imaging, Fujian Medical University
  4. Feng Pan; Imaging Department, First Affiliated Hospital of Fujian Medical University
  5. Yueming Li; Imaging Department, First Affiliated Hospital of Fujian Medical University
Objective:
As a noninvasive and reproducible technical tool, radiomics has shown great application prospects in the study of tumor pathology. Delta radiomics provides information on the temporal evolution of eigenvalues. We aimed to investigate the value of multimodality and multiregional radiomics as well as Delta radiomics in predicting MVI of HCC.

Materials and Methods:
A total of 161 patients with HCC confirmed by pathology at our hospital were retrospectively included and randomly divided into training and validation sets in the ratio of 7:3 (training set:validation set = 112:49). The tumor ROIs were outlined independently on T2WI and nonenhanced T1WI, arterial phase (AP), portal venous phase (PVP), delayed phase (DP), and hepatobiliary phase (HBP) images, using 3D Slicer software to outline the whole tumor area, the 10-mm peritumor area, and the whole tumor + 10-mm peritumor area of the lesion. The corresponding radiomics features were extracted using the pyradiomics package using whole-tumor features from different phases, including direct phase subtraction, normalized phase subtraction, and relative phase subtraction modalities, to calculate Delta radiomics features. In the training set, the maximum correlation and minimum redundancy algorithm and the recursive feature elimination algorithm were used to filter the radiomics features, the support vector machine algorithm was used to construct the radiomics models. The AUC values were used to evaluate the prediction performance of different models; the models were also validated in the validation set.

Results:
The AUC (95% CI) values of the PVP and HBP radiomics models based on the whole tumor + peritumor region were 0.828 (0.721–0.887) in the training set and 0.804 (0.675–0.934) in validation set. The AUC (95% CI) for the direct phase subtraction-based multiphase Delta radiomics model (portal phase – nonenhanced T1WI, delayed phase – nonenhanced T1WI, and PVP – AP) was 0.825 (0.747–0.904) and 0.771 (0.639–0.903) on the training and validation sets, respectively. The models based on whole tumor + peritumor region in PVP and HBP images had the best predictive efficacy among all radiomics models.

Conclusion:
We hope to provide noninvasive imaging biomarkers for the prediction of MVI through the combination of medicine and artificial intelligence techniques, further elucidate the potential biological significance behind imaging of radiomics features, and provide imaging basis for individualized clinical treatment of HCC.