How brains realize learning and perception is an essential question for both artificial intelligence and neuroscience communities. Although artificial neural networks (ANNs) are inspired from the computational principles in brains, the computing mechanism of ANN is far from the real brain. Thus, it is hard to use ANNs to explore the principles of learning and perception in the real brain. Dendritic neuron models are always used to simulate the signal integration process in the brain. Performing learning and perceptual tasks with dendritic neuron models help scientists better understand the learning mechanism in the real brain. However, current learning models on mainly focus on the simplified dendritic neurons, which is unable to model the entire signal processing mechanisms of real neurons. To solve this problem, we propose a biophysically detailed neural network of medium spiny neurons (MSNs). The detailed network model can perform image classification tasks after learning. Experimental results show that the detailed network can achieve high performance on the image classification task. In addition, the detailed network shows strong robustness under noise interference. With further analysis of the neurons in the network, we find that the detailed neurons show the stimulus selectivity property after learning, which is consistent with the findings in neuroscience research. This property indicates that our model is biologically-plausible, and implies that the stimulus selectivity is an essential property in perception learning.