An attention-based deep learning model for predicting microvascular invasion of hepatocellular carcinoma using an intra-voxel incoherent motion model of diffusion-weighted magnetic resonance imaging

The intra-voxel incoherent motion model of diffusion-weighted magnetic resonance imaging (IVIM-DWI) with a series of images with differentb-values has great potential as a tool for detecting, diagnosing, staging, and monitoring disease progression or the response to treatment. The current clinical tumour characterisation using IVIM-DWI is based on the parameter values derived from the IVIM model. On the one hand, the calculation accuracy of such parameter values is susceptible to deviations due to noise and motion; on the other hand, the performance of the parameter values is rather limited with respect to tumour characterisation. In this article, we propose a deep learning approach to directly extract spatiotemporal features from a series ofb-value images of IVIM-DWI using a deep learning network for lesion characterisation. Specifically, we introduce an attention mechanism to select dominant features from specificb-values, channels, and spatial areas of the multipleb-value images for better lesion characterisation. The experimental results for clinical hepatocellular carcinoma (HCC) when using IVIM-DWI demonstrate the superiority of the proposed deep learning model for predicting the microvascular invasion (MVI) of HCC. In addition, the ablation study reflects the effectiveness of the attention mechanism for improving MVI prediction. We believe that the proposed model may be a useful tool for the lesion characterisation of IVIM-DWI in clinical practice.