Deep learning model based on preoperative multimodality data predicting early recurrence in hepatocellular carcinoma

Background & Aims: To analyze the performance of multicenter preoperative multimodality data [magnetic resonance imaging (MRI) images and clinical characteristics] using deep learning method to predict early recurrence (≤ 2 years after surgery) in hepatocellular carcinoma (HCC) treated by radical surgery. Methods: Two MRI models based on multi-sequence MRI images were built using classical Resnet34 (mMRI_R model) and novel transformer (mMRI_T model) networks, respectively. Single-sequence MRI models (T1WI, T2WI, DWI and DCEI models) were built based on images of T1-weighted, T2-weighted, diffusion-weighted and dynamic contrast-enhanced MRI, respectively. For the optimal multi-sequence and single-sequence MRI models, we further proposed prediction models by integrating selected clinical characteristics. Results: A total of 489 patients were included (391 in training group and 98 in validation group). The mMRI_R model showed better discrimination than mMRI_T model in validation group (P = 0.037). Multi-sequence MRI images combined with selected clinical characteristics (mMRI+C_R model) achieved higher discrimination (mMRI_R, AUC = 0.826 vs. mMRI+C_R, AUC = 0.832; P = 0.002). T2WI model showed better discrimination than other single-sequence MRI models in validation group (P < 0.001). T2WI images combined with selected clinical characteristics (T2WI+C model) achieved higher discrimination (T2WI, AUC = 0.804 vs. T2WI+C, AUC = 0.811; P < 0.001). Conclusions: This deep learning model based on preoperative multi-sequence MRI images and selected clinical characteristics showed better discrimination for predicting early recurrence in HCC patients. Among single-sequence MRI models, model based on T2WI images and selected clinical characteristics showed better performance, which may be used as an alternative option.