Source-linked AI summary
Classification of breast cancer histology images using transfer learning
Sulaiman Vesal, Nishant Ravikumar, AmirAbbas Davari, Stephan Ellmann, Andreas Maier
TL;DR
Breast histology assessment is laborious and subjective, motivating automated classification across four tissue sub-types. The study normalizes images, extracts and augments patches, and fine-tunes ImageNet-pretrained Inception-V3 and ResNet50 networks; ResNet50 achieves 97.50% test accuracy versus 91.25% for Inception-V3.
Problem
Visual histopathology assessment is time-consuming and subjective, while automated classification of breast cancer sub-types remains challenging.
Method
The study applies stain normalization, overlapping patch extraction and augmentation, then fine-tunes ImageNet-pretrained Inception-V3 and ResNet50 networks on four breast histology classes.
Results
97.50% whole-image test accuracy was achieved by ResNet50, compared with 91.25% for Inception-V3.
Takeaways & Limitations
The study demonstrates the applicability of transfer learning for automatic breast histology classification with limited training data.
Abstract
from arXiv · showhide
Breast cancer is one of the leading causes of mortality in women. Early detection and treatment are imperative for improving survival rates, which have steadily increased in recent years as a result of more sophisticated computer-aided-diagnosis (CAD) systems. A critical component of breast cancer diagnosis relies on histopathology, a laborious and highly subjective process. Consequently, CAD systems are essential to reduce inter-rater variability and supplement the analyses conducted by specialists. In this paper, a transfer-learning based approach is proposed, for the task of breast histology image classification into four tissue sub-types, namely, normal, benign, \textit{in situ} carcinoma and invasive carcinoma. The histology images, provided as part of the BACH 2018 grand challenge, were first normalized to correct for color variations resulting from inconsistencies during slide preparation. Subsequently, image patches were extracted and used to fine-tune Google`s Inception-V3 and ResNet50 convolutional neural networks (CNNs), both pre-trained on the ImageNet database, enabling them to learn domain-specific features, necessary to classify the histology images. The ResNet50 network (based on residual learning) achieved a test classification accuracy of 97.50% for four classes, outperforming the Inception-V3 network which achieved an accuracy of 91.25%.
1 Introduction
Breast cancer diagnosis depends on histopathology, but visual assessment is time-consuming and subjective. Computer-aided diagnosis can reduce specialist workload and subjectivity while supporting classification of cancer sub-types.
- Histopathology is the gold standard for distinguishing benign from malignant tissue and in situ from invasive carcinoma.
- Visual assessment of tissue microstructure and nuclear organization is time-consuming and highly subjective.
- Computer-aided diagnosis systems are intended to improve diagnostic efficiency and reduce subjectivity in disease classification.
- Automated classification of histology images into cancer sub-types and metastasis detection in whole-slide images remain challenging tasks.
2 Methods
The study uses BACH 2018 breast histology images, stain normalization, overlapping patch extraction, augmentation, and ImageNet-pretrained CNN fine-tuning. Inception-V3 and ResNet50 are adapted to classify four tissue classes.
- Dataset: The BACH 2018 dataset contains H&E-stained images labeled as normal, benign, in situ carcinoma, or invasive carcinoma.Images were high-resolution and annotated by two medical experts; disagreements were discarded.
- Dataset: 400 images were split into training, validation, and test sets with 60, 20, and 20 samples per class, respectively.
- Stain Normalization: Stain normalization matches source-image color statistics to a target image in the decorrelated LAB color space.The method addresses color variation caused by scanners, materials, vendors, and staining protocols.
- Pre-processing: 512 × 512-pixel patches with 50% overlap were extracted to capture tissue architecture and localized nuclear organization from large images.This produced 35 patches per image.
- Pre-processing: Rotation and flipping augmentation produced 67,200 training and validation patches from the original 320 training images.Each patch generated five additional unique patches using horizontal and vertical flips and 90°, 180°, and 270° rotations.
- Pre-trained CNN Architectures: ImageNet-pretrained Inception-V3 and ResNet50 networks were fine-tuned to learn domain-specific features for breast histology classification.The workflow applies preprocessing before the two network architectures.
3 Results
ResNet50 outperformed Inception-V3 on patch-wise and whole-image accuracy for validation and test data. ROC analysis also reported higher overall specificity and sensitivity for ResNet50.
- Classification Accuracy: 97.50% whole-image test accuracy was achieved by ResNet50, compared with 91.25% for Inception-V3.
- Classification Accuracy: 94.50% patch-wise test accuracy was achieved by ResNet50, compared with 86.57% for Inception-V3.
- Classification Accuracy: ResNet50 consistently outperformed Inception-V3 on patch-wise and image-wise accuracy for both validation and test data.
- ROC Analysis: Overall specificity and sensitivity were approximately 99.9% for ResNet50 and 98% for Inception-V3.ROC curves and class-specific AUCs were computed on the unseen test set.
4 Conclusions
The study presents transfer learning for four-class classification of H&E-stained breast histology images using ImageNet-pretrained Inception-V3 and ResNet50. ResNet50 consistently outperformed Inception-V3, supporting the applicability of transfer learning with limited training data.
- The approach classifies normal, benign, in situ carcinoma, and invasive carcinoma images using ImageNet-pretrained CNN architectures.
- The networks were trained on 60% of the dataset, validated on 20%, and evaluated on the remaining 20% of images.
- ResNet50 consistently outperformed Inception-V3 in classification accuracy.
- The results demonstrate transfer learning’s applicability to breast histology classification in the presence of limited training data.