Source-linked AI summary
Deep Convolutional Neural Networks for Breast Cancer Histology Image Analysis
Alexander Rakhlin, Alexey Shvets, Vladimir Iglovikov, Alexandr A. Kalinin
TL;DR
Breast histology assessment is tedious and subjective, with limited data challenging deep-learning training. The paper combines pretrained deep CNN features, augmentation, and gradient-boosted trees, reporting 87.2% accuracy for four-class classification and 93.8% accuracy with AUC 0.973 for carcinoma detection.
Problem
Pathologists’ visual assessment of breast histology is tedious and subjective, while the small four-class dataset challenges deep-learning training.
Method
The approach extracts multiscale features from ImageNet-pretrained deep CNNs, applies strong augmentation, and uses gradient-boosted trees for classification.
Results
87.2±2.6% accuracy was achieved for four-class classification, while carcinoma detection achieved 93.8±2.3% accuracy, AUC 0.973, and 96.5% sensitivity with 88.0% specificity at the high-sensitivity setpoint.
Takeaways & Limitations
The authors report that the method outperforms similar automated breast-image analysis solutions and literature methods in this small-data setting.
Abstract
from arXiv · showhide
Breast cancer is one of the main causes of cancer death worldwide. Early diagnostics significantly increases the chances of correct treatment and survival, but this process is tedious and often leads to a disagreement between pathologists. Computer-aided diagnosis systems showed potential for improving the diagnostic accuracy. In this work, we develop the computational approach based on deep convolution neural networks for breast cancer histology image classification. Hematoxylin and eosin stained breast histology microscopy image dataset is provided as a part of the ICIAR 2018 Grand Challenge on Breast Cancer Histology Images. Our approach utilizes several deep neural network architectures and gradient boosted trees classifier. For 4-class classification task, we report 87.2% accuracy. For 2-class classification task to detect carcinomas we report 93.8% accuracy, AUC 97.3%, and sensitivity/specificity 96.5/88.0% at the high-sensitivity operating point. To our knowledge, this approach outperforms other common methods in automated histopathological image classification. The source code for our approach is made publicly available at https://github.com/alexander-rakhlin/ICIAR2018
1 Introduction
Breast histology diagnosis distinguishes normal, benign, in situ, and invasive tissue, but visual assessment is tedious and subjective. The paper therefore applies deep CNN features and gradient boosted trees to breast cancer type classification.
- Histopathology task: Breast histopathology distinguishes normal tissue, benign lesions, and carcinomas, including in situ and invasive carcinoma.Invasive carcinoma involves cancer cells breaking through duct walls into surrounding tissue.
- Diagnostic motivation: Pathologists traditionally assess morphology visually, but the process is tedious and produces inter-observer variation.The variation occurs even among senior pathologists.
- Diagnostic motivation: Computer-aided diagnosis is motivated by goals of improving accuracy, reducing human error, increasing agreement, and improving reproducibility.
- Related work: Deep learning approaches have outperformed conventional machine learning methods on many image-analysis tasks while automating end-to-end processing.CNNs have also been used in several medical-imaging applications.
- Paper contribution: The proposed approach uses deep CNNs for feature extraction and gradient boosted trees for breast histology image classification.The authors report that it outperforms other similar solutions.
2 Methods
The method addresses a small, balanced four-class H&E image dataset using pretrained CNN feature extraction, augmentation, multiscale descriptors, and LightGBM classification. Cross-validation and model fusion produce predictions while preventing image-level information leakage.
- 2.1 Dataset: The dataset contains 400 H&E-stained images labeled as normal, benign, in situ carcinoma, or invasive carcinoma.The four classes are balanced, and each image was annotated by two medical experts.
- 2.2 Approach overview: The 400-image dataset creates a significant challenge for training deep learning models.Very deep CNNs contain millions of parameters and can overfit small datasets.
- 2.2 Approach overview: Pretrained CNNs extract low-dimensional descriptors instead of being fine-tuned directly on the small dataset.The descriptors have dimensionality 1408 or 2048, reducing overfitting risk in supervised learning.
- 2.3 Data pre-processing and augmentation: H&E color augmentation, multiple crop sizes, and three CNN encoders expand each image into multiple descriptors for pooled representation.The preprocessing pipeline standardizes stain amounts and uses random H&E perturbations.
- 2.4 Feature extraction: ResNet-50, InceptionV3, and VGG-16 provide deep features, with fully connected layers removed and convolutional outputs globally pooled.The resulting vectors support feature extraction from images of arbitrary size.
- 2.5 Training: The training procedure uses stratified folds, keeps descriptors from each image within one fold, and trains 600 LightGBM models across seeds, scales, crops, and encoders.Test predictions are averaged across augmentations and models before selecting the maximum class probability.
3 Results
Across the evaluated tasks, the approach achieved strong breast histology classification performance, including 87.2% accuracy for four classes and 93.8% accuracy with AUC 0.973 for carcinoma detection.
- 2-class classification: 96.5% sensitivity and 88.0% specificity were obtained at the high-sensitivity setpoint of 0.33 for detecting carcinomas.At setpoint 0.50, sensitivity was 93.0% and specificity was 94.5%.
- 2-class classification: Only 9 in situ and 5 invasive carcinoma cases were missed among 200 carcinoma cases.This result corresponds to the reported confusion matrix analysis.
- 4-class classification: 87.2±2.6% accuracy was achieved for four-class classification averaged across 10 cross-validation folds.The four classes were evaluated using stratified cross-validation.
- Model fusion: The fused model accuracy was 4-5% higher than that of any individual constituent model.The ensemble’s standard deviation across folds was twice as low as the average standard deviation of individual models.
4 Conclusions
The paper presents a method for classifying H&E-stained breast histology images with very small training data by combining deep convolutional features, augmentation, and gradient boosting. The authors report results superior to automated breast cancer image analysis methods in the literature.
- Method: The proposed method targets H&E-stained breast cancer image classification when only a few hundred training samples are available.The approach is designed for the very small-data setting.
- Method: Deep convolutional features from ImageNet-pretrained CNNs are extracted at multiple scales and classified with gradient boosting.Strong data augmentation is used to increase classifier robustness.
- Conclusion: The authors report that their results are superior to automated breast cancer image analysis reported in the literature.This is presented as the authors’ assessment of the reported results.