Abstract: Due to the high heterogeneity of breast cancer, different molecular subtypes of breast cancer have great differences in the treatment and diagnosis. Therefore, it is of important research significance to improve the accuracy of the diagnosis of breast cancer subtypes, thus to further reduce the misdiagnosis rate of breast cancer and avoid overtreatment of breast cancer. The molecular subtypes of Luminal A and non-Luminal A of breast cancer were classified based on deep learning algorithm for MRI breast cancer medical images in TCIA database. In order to compare the superiority of deep learning algorithm in classification of breast cancer molecular subtypes, gene expression data of TCGA database was used to classify breast cancer molecular subtypes. In the same breast cancer MRI image database, the classification of breast cancer molecular subtypes was studied based on traditional machine learning algorithm. Comparative studies between a variety of deep learning models were conducted in the processing of MRI images with deep learning algorithm. Fine-tuning VGG16 network and frozen convolution layer were implemented. An improved VGG16 network model was proposed, and a densenet network module was added, namely VGG16+densenet（4） model. The accuracy of the improved VGG16 network model was 0. 96, and the AUC was 0. 97. The classification accuracy of gene expression data was 0. 73, and the AUC was 0. 79. The classification accuracy and AUC of traditional machine learning reached 0. 80 and 0. 87 respectively. The experimental results show that the proposed VGG16+densenet（4） model improves the accuracy of the molecular subtypes of breast cancer and has better classification effect.