Transfer Learning and One Class Classification - A Combined Approach for Tumor Classification

Abstract

Deep learning models have extended its application in computer aided diagnosis of various medical complications. Identification of tumors from the images obtained from Magnetic Resonance Imaging (MRI) is one among them. But, in certain situations where the availability of dataset, in specific, the number of observations in a particular class, is very low than the other class, techniques such as one-class classification has to be incurred. This work combines the concept of transfer learning and one-class classification. The best pre-trained CNN which is capable of classifying the MRI images with tumors and without tumors is identified and is used for feature extraction. The features are extracted from a dataset with 465 positive images and 46 negative images. The extracted features are given as input to the one-class classifiers. The pre-trained models compared are VGG19, Resnet50 and Densenet121. VGG19 shows the best performance and hence used for feature extraction. The one-class classifiers compared are one-class support vector machine and isolation forest. One-class support vector machine performs better than the isolation forest algorithm.