Computer aided diagnosis of degenerative intervertebral disc disease is a challenging task which has been targeted many times by computer vision and image processing community. This paper proposes a deep network approach for the diagnosis of degenerative intervertebral disc disease. Different from the classical deep networks, our system uses non-linear filters between the network layers that introduce domain dependent information into the network training for a faster training with lesser amount of data. The proposed system takes advantage of the unsupervised feature extraction with deep networks while requiring only a small amount of training data, which is a major problem for medical image analysis where obtaining large amounts of patient data is very difficult. The method is validated on a dataset containing 102 lumbar MR images. State-of-the-art hand-crafted feature extraction algorithms are compared with the unsupervisedly learned features and the proposed method outperforms the hand-crafted features.