BREAST HISTOPATHOLOGY IMAGES MULTI- CLASSIFICATION WITH ENSEMBLES OF DEEP CONVOLUTIONAL NEURAL NETWORKS

Abstract

Medical disease diagnosis has benefitted greatly from advancements in computing methods that seek patterns in medical data for diagnosis and predictions of diseases. Breast cancers have constituted a major health challenge as a leading cause of mortality in women. Classification and analysis of breast histopathology images digitally provides a means for computerized-clinical diagnosis of breast cancers. A literature review of some works providing models that can classify histopathology images and thereby diagnose breast cancers was carried out. Three separate models based on the ensembles of deep Convolutional Neural Networks (CNNs) were developed. The ensemble models extended the performance of existing CNNs by combining AlexNet, VGGNet and ResNet pre-trained models using majority voting, averaging and stacking ensemble rules. Furthermore, a model for classification of breast histopathology images was developed called the SaduNet model. This model was developed with few number of convolutional neural network layers to reduce computational cost in terms of memory and improve computation time. All the four models were trained with histopathology images dataset collected at the Federal Teaching Hospital, Ido Ekiti, Ekiti State, Nigeria, with ethical clearance to ensure that the research is applicable locally. Different learning parameters were used in the different convolutional neural networks developed to ensure that they obtain high performances on the histopathology images classification task. The comparative analysis performed showed that the developed ensemble models performed as well as those found in literature judging by the accuracies achieved. The ensemble models also performed better in terms of sensitivity and predictability than the individual base models. This is shown in the high prediction and recall values obtained by the ensemble models. When tested with the Federal Teaching Hospital breast histopathology data, the averaging ensemble has accuracy of 99.47% while the majority voting ensemble has accuracy of 9.30% and stacking ensemble model has accuracy of 97.86%. The SaduNet model achieved an accuracy of 75.89% when tested with the FTH dataset. Also, when tested with the BreakHis dataset, the averaging ensemble has accuracy of 98.32% while the majority voting and stacking ensemble models has accuracies of 97.90% respectively. The SaduNet model also achieved an accuracy of 79.2%. SaduNet has 43,064,448 trainable parameters. Thus, SaduNet model provides less computation cost because it has less number of trainable parameters in its neural network. This research provides an application of machine learning to breast cancer diagnosis and contributes to the knowledge on classification of breast histopathology images with deep convolutional neural networks.