Detection of GH Pituitary Tumors Based on MNF

In recent years, more and more machine learning algorithms have been applied to the field of medicine, but the medical data is often very few because of the high cost of data acquisition. Although machine learning has recently made significant breakthroughs in many areas, learning from small datasets remains a challenge for machine learning. In contrast, many metric learning algorithms allow novel examples to be rapidly assimilated, rather than suffering from serious forgetting. This paper aims at the detection of GH pituitary tumors under a small dataset, and inspired by the attention mechanism, the Metric Learning Based on Neural Features (MNF) is proposed by constructing a kernel that combines metric learning with deep convolutional neural networks. In addition, this paper also uses Long Short-Term Memory (LSTM) to further optimize the MNF. Finally, this paper also uses transfer learning to train the convolutional neural networks and use data augmentation methods based on ROI to improve the final results. Experiments show that MNF can be used under the small dataset, and get an accuracy of 92.3% in the detection of GH pituitary tumors.