Classification of Breast Tumours Based on Histopathology Images Using Deep Features and Ensemble of Gradient Boosting Methods

TL;DR

This paper presents a deep learning-based CAD system for breast tumor classification using histopathology images, employing transfer learning with Inception-ResNet-v2 for feature extraction and an ensemble of gradient boosting methods for classification, achieving high accuracy on the BreakHis dataset.

Contribution

It introduces a novel combination of deep transfer learning and ensemble gradient boosting methods for improved breast tumor classification accuracy.

Findings

Achieved over 96% accuracy across multiple magnifications.

Inception-ResNet-v2 outperformed other CNNs in feature extraction.

Ensemble of CatBoost, XGBoost, and LightGBM improved classification results.

Abstract

Breast cancer is the most common cancer among women worldwide. Early-stage diagnosis of breast cancer can significantly improve the efficiency of treatment. Computer-aided diagnosis (CAD) systems are widely adopted in this issue due to their reliability, accuracy and affordability. There are different imaging techniques for a breast cancer diagnosis; one of the most accurate ones is histopathology which is used in this paper. Deep feature transfer learning is used as the main idea of the proposed CAD system's feature extractor. Although 16 different pre-trained networks have been tested in this study, our main focus is on the classification phase. The Inception-ResNet-v2 which has both residual and inception networks profits together has shown the best feature extraction capability in the case of breast cancer histopathology images among all tested CNNs. In the classification phase, the ensemble of CatBoost, XGBoost and LightGBM has provided the best average accuracy. The BreakHis dataset was used to evaluate the proposed method. BreakHis contains 7909 histopathology images (2,480 benign and 5,429 malignant) in four magnification factors. The proposed method's accuracy (IRv2-CXL) using 70% of BreakHis dataset as training data in 40x, 100x, 200x and 400x magnification is 96.82%, 95.84%, 97.01% and 96.15%, respectively. Most studies on automated breast cancer detection have focused on feature extraction, which made us attend to the classification phase. IRv2-CXL has shown better or comparable results in all magnifications due to using the soft voting ensemble method which could combine the advantages of CatBoost, XGBoost and LightGBM together.