SegEIR-Net: A Robust Histopathology Image Analysis Framework for Accurate Breast Cancer Classification

Abstract

Background: Breast Cancer (BC) is a significant threat affecting women globally. An accurate and reliable disease classification method is required to get an early diagnosis. However, existing approaches lack accurate and robust classification.

Objective: This study aims to design a model to classify BC Histopathology images accurately by leveraging segmentation techniques.

Methods: This work proposes a combined segmentation and classification approach for classifying BC using histopathology images to address these issues. Chan-Vese algorithm is used for segmentation to accurately delineate regions of interest within the histopathology images, followed by the proposed SegEIR-Net (Segmentation using EfficientNet, InceptionNet, and ResNet) for classification. Bilateral Filtering is also employed for noise reduction. The proposed model uses three significant networks, ResNet, InceptionNet, and EfficientNet, concatenates the outputs from each block followed by Dense and Dropout layers. The model is trained on the breakHis dataset for four different magnifications and tested on BACH (BreAst Cancer Histology) and UCSB (University of California, Santa Barbara) datasets.

Results: SegEIR-Net performs better than the existing State-of-the-Art (SOTA) methods in terms of accuracy on all three datasets, proving the robustness of the proposed model. The accuracy achieved on breakHis dataset are 98.66%, 98.39%, 97.52%, 95.22% on different magnifications, and 93.33% and 96.55% on BACH and UCSB datasets.

Conclusion: These performance results indicate the robustness of the proposed SegEIR-Net framework in accurately classifying BC from histopathology images.

Keywords: Breast Cancer (BC); Deep Learning; Histopathology Classification; InceptionNet; ResNet; and EfficientNet.