Histopathological Image Classification: Defying Deep Architectures on Complex Data

Abstract

Automatic analysis of medical images is a challenging research which requires both the skill of a pathologist and computer vision knowledge to develop efficient systems. In this work, we have taken up the task of classifying different types of cell nuclei in histopathological Colon Cancer Images. We aim to show the relevance and effect of a complex histopathological dataset on the performance of current deep learning architectures. We have experimented with pre-trained (on ImageNet) AlexNet, VGG16, and VGG19 architectures and applied transfer learning approach to train these architectures. On the basis of the results obtained on the Histopathological image dataset, while using fine tuned AlexNet, VGG16, and VGG19 architectures; the suitability of using pure architectures is somehow questionable and these state of the art algorithms straightaway cannot be used for the sophisticated classification of very complex cancer tissue dataset. Comparative evaluation of the above state of the art methods have been done and the possibility of devising hybrid deep architectures is investigated thereof.