Histopathological Imaging Classification of Breast Tissue for Cancer Diagnosis Support Using Deep Learning Models

TL;DR

This paper presents a deep learning approach using EfficientNet to classify breast histopathology images into four categories, achieving high accuracy and aiding cancer diagnosis.

Contribution

It introduces a data augmentation method with overlapping patches and applies EfficientNet for accurate breast cancer histopathology classification.

Findings

98% training accuracy

93% evaluation accuracy

Effective for cancer diagnosis support

Abstract

According to some medical imaging techniques, breast histopathology images called Hematoxylin and Eosin are considered as the gold standard for cancer diagnoses. Based on the idea of dividing the pathologic image (WSI) into multiple patches, we used the window [512,512] sliding from left to right and sliding from top to bottom, each sliding step overlapping by 50% to augmented data on a dataset of 400 images which were gathered from the ICIAR 2018 Grand Challenge. Then use the EffficientNet model to classify and identify the histopathological images of breast cancer into 4 types: Normal, Benign, Carcinoma, Invasive Carcinoma. The EffficientNet model is a recently developed model that uniformly scales the width, depth, and resolution of the network with a set of fixed scaling factors that are well suited for training images with high resolution. And the results of this model give a rather competitive classification efficiency, achieving 98% accuracy on the training set and 93% on the evaluation set.