DEEP ENSEMBLE TRANSFER LEARNING FOR DETECTING BREAST CANCER IN HISTOPATHOLOGICAL IMAGES

Abstract

Among other cancer type, breast cancer is the leading cause of death worldwide. The traditional approach in detecting breast cancer malignancy relied on rigorous analysis, making the whole process prone to diagnostic error. This study proposes a deep learning solution to solve the problem using deep ensemble convolutional neural network (CNN) models constructed from single and smaller ensemble models. The utilized single models are ResNet50V2, InceptionResNet50V2, DenseNet201, EfficientNetB4, EfficientNetV2S, and Xception. Smaller ensemble combinations are also made from the single models. The deep ensemble models composed of ResNet50V2-EfficientNetV2S-DenseNet201, EfficientNetB4-EfficientNetV2S-Xception, EfficientNetB4-EfficientNetV2S, DenseNet201-EfficientNetB4, and ResNet50V2-DenseNet201. These models are trained using histopathological images acquired from Hasanuddin University Hospital and BreakHis with 400x magnification. Despite the data imbalance, the deep ensemble successfully obtained a 0.94 ROC-AUC score with a 0.97 average precision (AP) score, showing its capability to distinguish breast cancer malignancy from histopathological images. Further analysis revealed some distinctive patterns in the image that make the images easily classified by the deep ensemble model. This study has demonstrated that the deep ensemble CNN model constructed from smaller ensemble CNN models yields remarkable results in breast cancer detection.

Access to Document

Other files and links

• Link to publication in Scopus
• Open Access Version Available

Fingerprint