[Paper Review] Classification of Histopathological Biopsy Images Using Ensemble of Deep Learning Networks

Sara Hosseinzadeh Kassani, Peyman Hosseinzadeh Kassani+3 more|arXiv (Cornell University)|Sep 26, 2019

AI in cancer detectionComputer Science38 references60 citations

TL;DR

This paper proposes a three-path ensemble of pre-trained CNNs (VGG19, MobileNet, DenseNet201) with transfer learning to binary classify breast histopathology images, achieving state-of-the-art results on four public datasets.

ABSTRACT

Breast cancer is one of the leading causes of death across the world in women. Early diagnosis of this type of cancer is critical for treatment and patient care. Computer-aided detection (CAD) systems using convolutional neural networks (CNN) could assist in the classification of abnormalities. In this study, we proposed an ensemble deep learning-based approach for automatic binary classification of breast histology images. The proposed ensemble model adapts three pre-trained CNNs, namely VGG19, MobileNet, and DenseNet. The ensemble model is used for the feature representation and extraction steps. The extracted features are then fed into a multi-layer perceptron classifier to carry out the classification task. Various pre-processing and CNN tuning techniques such as stain-normalization, data augmentation, hyperparameter tuning, and fine-tuning are used to train the model. The proposed method is validated on four publicly available benchmark datasets, i.e., ICIAR, BreakHis, PatchCamelyon, and Bioimaging. The proposed multi-model ensemble method obtains better predictions than single classifiers and machine learning algorithms with accuracies of 98.13%, 95.00%, 94.64% and 83.10% for BreakHis, ICIAR, PatchCamelyon and Bioimaging datasets, respectively.

Motivation & Objective

• Address the challenge of heterogeneous histopathology image data from multiple sources.
• Develop a robust binary classifier for benign vs malignant breast histology images without handcrafted features.
• Leverage transfer learning and data augmentation to improve generalization across datasets.
• Evaluate the ensemble model against single CNNs and traditional machine learning methods on public benchmarks.

Proposed method

• Three-path ensemble architecture combining VGG19, MobileNetV2, and DenseNet201 for feature extraction.
• Transfer learning from ImageNet-weighted CNNs, with fine-tuning to output two classes.
• Flatten and concatenate the final layers to form a multi-view feature vector fed into a multi-layer perceptron classifier.
• Pre-processing includes stain normalization (Macenko), image normalization, and data augmentation (flips, rotations, zoom, etc.).
• Training specifics: images resized to 224x224, batch size 32, 1000 epochs, Adam optimizer, dropout 0.5, 256-hidden neuron FC layer.

Experimental results

Research questions

• RQ1 Can a three-path CNN ensemble improve binary breast histopathology classification over single CNNs and traditional ML methods across multiple datasets?
• RQ2 How does transfer learning with stain normalization and augmentation affect generalization to heterogeneous datasets (BreakHis, ICIAR, PatchCamelyon, Bioimaging)?
• RQ3 What are the comparative accuracies and other metrics of the ensemble versus individual architectures on public benchmarks?

Key findings

• On BreakHis, the ensemble achieved 98.13% accuracy, 98.75% precision, 98.54% recall, and 98.64% F-score.
• On PatchCamelyon*, the ensemble achieved 94.64% accuracy, 95.70% precision, 95.27% recall, and 95.50% F-score.
• On ICIAR, the ensemble achieved 95.00% accuracy, 95.91% precision, 94.00% recall, and 94.94% F-score.
• On Bioimaging, the ensemble achieved 83.10% accuracy, 92.60% precision, 71.42% recall, and 80.64% F-score.
• Single CNNs (VGG19, MobileNetV2, DenseNet201) and other state-of-the-art CNNs generally underperformed the ensemble on BreakHis, ICIAR, and PatchCamelyon* datasets; Bioimaging remained challenging with overall lower performance.
• Compared with several machine learning models (Decision Tree, Random Forest, XGBoost, AdaBoost, Bagging), the ensemble showed superior performance on most datasets, though some literature methods achieved higher accuracy on specific datasets (e.g., Pratiher & Chattoraj 2019 on BreakHis).