Comparing Deep Learners with Variability Grading for Cancer Detection on Limited Histopathology Dataset

Abstract

State-of-the-art deep convolution neural networks (CNN) can be applied to various domains, including the grading of cancers in histopathology images, and are most promising approaches. However, it is well-known that CNNs require huge amounts of tagged images and resources to train and work well, and some prior works on cancer grading also achieved top accuracy by analyzing how cancer affects structures, such as cells, in terms of variability of characteristics. The aim of this work is to compare CNN-based classification of medical images with automated analysis of multiple structures. This is done experimentally, by implementing the alternatives and comparing classification accuracy on a public cancer grading dataset. The results show that a well-designed automated analysis of structures improved accuracy by 4% when compared with the best CNN result, showing that it is worth to study further and establish procedures based on that analysis.