On evaluating CNN representations for low resource medical image classification

A few examples with custom CNN design include acoustic in several machine learning tasks such as image classification, modeling for low resource languages [8], object and action classification object tracking, and keyword spotting. However, given that [9] and remote sensing [10]. On the other hand, medical image they contain a large number of parameters, their direct applicability classification [11] requires assignment of medical images (drawn into low resource tasks is not straightforward. In this work, we from real world patients) to a medical landmark, phenomenon or a experiment with an application of CNN models to gastrointestinal disease and often, obtaining large amounts of training data can be landmark classification with only a few thousands of training samples challenging. A few approaches for medical image classification include through transfer learning. As in a standard transfer learning the use of decision trees [12], k-nearest-neighbors [13] and approach, we train CNNs on a large external corpus, followed by support vector machines [14].