Comparison Of Deep Learning-Based Recognition Techniques For Medical And Biomedical Images

The recognition and classification of medical and biomedical images typically suffer from the problem of a low number of annotated samples. This comes along with the problem of efficient training of the current deep learning frameworks. Therefore, many researchers opt for various techniques which could substitute the traditional training of convolutional neural networks (CNN) from scratch. In this article, we are comparing multiple of these methods, including transfer learning and using the CNNs as feature extractors. The paper contains results on two datasets with different modalities and three different CNN architectures. We demonstrate the high effectiveness of transfer learning and suggest that, in some cases, it is worth retraining more layers at the end of the network for achieving higher accuracy.