OBJECTIVE : To propose an automatic approach based on a convolutional neural network (CNN) to evaluate the quality of T2-weighted liver magnetic resonance (MR) images as nondiagnostic (ND) or diagnostic (D). METHODS : We included 150 T2-weighted liver MR imaging examinations in this retrospective study. Each slice of liver image was annotated with a label D or ND by two radiologists with seven and six years of experience, respectively. Additionally, the radiologists segmented the liver region manually as the ground truth for liver segmentation. A CNN was trained to segment the liver region and another CNN was used to classify the qualities of patches extracted from the liver region. The quality of an image was obtained from the percentage of nondiagnostic patches in all liver patches in the image. Treating nondiagnostic as positive, the accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), area under the receiver operating characteristic curve (AUC), and confusion matrix were used to evaluate our model. A Mann-Whitney U test was performed with the statistical significance set at 0.05. RESULTS : Our model achieved good performance with an accuracy of 88.3 %, sensitivity of 86.0 %, specificity of 89.4 %, PPV of 78.6 %, NPV of 93.4 %, and AUC of 0.911 (95 % confidence interval: 0.882-0.939, p < 0.05). The confusion matrix of our model indicated good concordance with that of the radiologists. CONCLUSIONS : The proposed two-step patch-based model achieved excellent performance when assessing the quality of liver MR images.