Exploring CNNs: an application study on nuclei recognition task in colon cancer histology images

In this work we explore the recent advances in the field of Convolutional Neural Network (CNN), with particular interest to the task of image classification. Moreover, we explore a new neural network algorithm, called ladder network, which enables the semi-supervised framework on pre-existing neural networks. These techniques were applied to a task of nuclei classification in routine colon cancer histology images. Specifically, starting from an existing CNN developed for this purpose, we improve its performances utilizing a better data augmentation, a more efficient initialization of the network and adding the batch normalization layer. These improvements were made to achieve a state-of-the-art architecture which could be compatible with the ladder network algorithm. A specific custom version of the ladder network algorithm was implemented in our CNN in order to use the amount of data without a label presented with the used database. However we observed a deterioration of the performances using the unlabeled examples of this database, probably due to a distribution bias in them compared to the labeled ones. Even without using of the semi-supervised framework, the ladder algorithm allows to obtain a better representation in the CNN which leads to a dramatic performance improvement of the starting CNN algorithm. We reach this result only with a little increase in complexity of the final model, working specifically on the training process of the algorithm.