The diagnosis and follow-up of focal liver lesions have an important place in radiology practice and in planning the treatment of patients. Lesions detected in the liver can be benign or malign. While benign lesions do not require any treatment, some treatments and surgical operations may be required for malign lesions. Magnetic resonance imaging provides some advantages over other imaging modalities in the detection and characterization of focal liver lesions with its superior soft tissue contrast. Additionally, different phases help make a clear diagnosis of different contrast agent retention properties in magnetic resonance imaging. This study aims to classify focal liver lesions based on convolutional neural networks by fusing magnetic resonance liver images obtained in pre-contrast, venous, arterial, and delayed phases. Magnetic resonance imaging data were obtained from Selcuk University, Faculty of Medicine, Department of Radiology in Turkey. The experiments were performed using 460 magnetic resonance images in four phases of 115 patients. Two experiments were conducted. Two-dimensional discrete wavelet transform was used to fuse the phases in both experiments. In the first experiment, the best model was determined using the original data, different number of convolution layers and different activation functions. In the second experiment, the best-found model was used. Additionally, the number of data was increased using data augmentation methods in this experiment. The results were compared with other state-of-the art methods and the superiority of the proposed method was proved. As a result of the classification, 96.66% accuracy, 86.67% sensitivity and 98.76% specificity rates were obtained. When the results are examined, CNN efficiency increases by fusing MR liver images taken in different phases.