Approach To Classifying Data With Highly Localized Unmarked Features Using Neural Networks

To face the increasing demand of quality healthcare, cutting-edge automation technology is being applied in demanding areas such as medical imaging. This paper proposes a novel approach to classification problems on datasets with sparse highly localized features. It is based on the use of a saliency map in the amplification of features. Unlike previous efforts, this approach does not use any prior information about feature localization. We present an experimental study based on the Diabetic Retinopathy classification problem, in which our method has shown to achieve an over 20%-higher accuracy in solving a two-class Diabetic Retinopathy classification problem than a naive approach based solely on residual neural networks. The dataset consists of 35,120 images of various qualities, inconsistent resolutions, and aspect ratios.