IVIM using convolutional neural networks predicts microvascular invasion in HCC

Objectives The study aimed to investigate the diagnostic performance of intravoxel incoherent motion (IVIM) diffusion-weighted magnetic resonance imaging for prediction of microvascular invasion (MVI) in hepatocellular carcinoma (HCC) using convolutional neural networks (CNNs). Methods This retrospective study included 114 patients with pathologically confirmed HCC from December 2014 to August 2021. All patients underwent MRI examination including IVIM sequence with 9 b -values preoperatively. First, 9 b -value images were superimposed in the channel dimension, and a b -value volume with a shape of 32 × 32 × 9 dimension was obtained. Secondly, an image resampling method was performed for data augmentation to generate more samples for training. Finally, deep features to predict MVI in HCC were directly derived from a b -value volume based on the CNN. Moreover, a deep learning model based on parameter maps and a fusion model combined with deep features of IVIM, clinical characteristics, and IVIM parameters were also constructed. Receiver operating characteristic (ROC) curve analysis was performed to assess the diagnostic performance for MVI prediction in HCC. Results Deep features directly extracted from IVIM-DWI (0.810 (range 0.760, 0.829)) using CNN yielded better performance for prediction of MVI than those from IVIM parameter maps (0.590 (range 0.555, 0.643)). Furthermore, the performance of the fusion model combined with deep features of IVIM-DWI, clinical features (α-fetoprotein (AFP) level and tumor size), and apparent diffusion coefficient (ADC) (0.829 (range 0.776, 0.848)) was slightly improved. Conclusions Deep learning with CNN based on IVIM-DWI can be conducive to preoperative prediction of MVI in patients with HCC. Key Points • Deep learning assessment of IVIM data for prediction of MVI in HCC can overcome the unstable and low performance of IVIM parameters. • Deep learning model based on IVIM performs better than parameter values, clinical features, and deep learning model based on parameter maps. • The fusion model combined with deep features of IVIM, clinical characteristics, and ADC yields better performance for prediction of MVI than the model only based on IVIM.