Curriculum Incremental Deep Learning on BreakHis DataSet.

This paper examines methodological aspects of the training procedure of neural networks for medical image classification. The research question concerns the conjecture that: feeding a network with datasets of increasing magnification leverages high-level knowledge and helps the network to better classify. This study confirms this hypothesis by an experiment carried out on a dataset of breast cancer histopathological images. Results are presented that underline the importance of the order in which data is introduced to the neural network during the training phase. Extensive experiments done on the BreakHis dataset demonstrate that curriculum incremental learning reaches 98.76% accuracy for binary classification while the best state of the art approach only reaches 96.78.%. Concerning multi-class classification, curriculum incremental learning reaches 95.93% while the state of the art approaches only reaches 95.49%. Moreover both the computational time and the stabilization time of the learning process of the incremental curriculum learning approach are reduced (respectively by 6% and by more than 20%) wrt a non curriculum learning approach.