Detecting outlier and poor quality medical images with an ensemble-based deep learning system

There are numerous reasons why inappropriate data can occur in a database. It is essential to detect and eliminate these elements for getting accurate results and conclusions. The process of filtering anomalies generated in the database is called outlier detection. The outliers, that are extreme values deviating from other observations on data, indicate a variability in a measurement, experimental errors or a novelty. In this paper, an ensemble-based outlier detection method is presented, where the members of the ensemble are convolutional neural networks (CNNs) combined with a Support Vector Machine (SVM) classifier. Ensemble-based methods are highly popular approaches that increase the accuracy of a decision by aggregating the opinions of individual voters. From hybrid models constructed by the neural networks and the SVM classifier, we have created an ensemble system that makes decision based on the conception of majority voting and results in very accurate outlier filtering. It actually serves as a pre-filter that can be integrated into a more exhaustive image analysis process with rejecting images that fall outside the domain or have poor quality. We have tested the performance of the proposed method for filtering databases consisting of retinal and skin lesion images, respectively. Our results show that the proposed ensemble system improved the effectiveness of the member hybrid CNN-SVMs in the outlier detection for retinal images and for skin lesions, as well.