Objective: To improve the accuracy of the prediction of primary liver cancer by radiomics employing non invasive hyperspectral imaging technology. Methods: This retrospective study was approved by the IRB committee of our institution, and the rule of informed consent was put forward. This paper analyzes the hyperspectral images (HSI) of patients with primary liver cancer and finds that the spectral reflectance curve of tumor and non-tumor locations are significantly different at different wavelengths. By using the ISODATA algorithm to predict tumor, it is found that the result is obvious. It is speculated that the use of HSI to select computed tomography (CT) images can improve the accuracy of radiomics in the prediction of primary liver cancer. Therefore, we employed HSI in combination with radiomic features to complete this task. We use the attention mechanism to concentrate computing resources into highly correlated features and use these features to predict tumor. Results: Using the hyperspectral-attention mechanism feature selection method, we were able to achieve an AUC of 0.96. Our study shows that HSI and radiomic features could aid in the prediction of liver cancer.