Source-linked AI summary
Two-Stage Convolutional Neural Network for Breast Cancer Histology Image Classification
Kamyar Nazeri, Azad Aminpour, Mehran Ebrahimi
TL;DR
The paper addresses four-class breast histology image classification under the memory constraints imposed by large microscopy images. It uses separate patch-wise and image-wise CNNs to combine local and global information. On the BACH validation set, the method achieves 95% accuracy and is reported as suitable for practical settings because of its small weights and millisecond-order inference.
Problem
The paper asks how to classify large breast histology microscopy images into normal, benign, in situ carcinoma, and invasive carcinoma despite hardware limitations.
Method
The method pre-trains a patch-wise CNN to extract local feature maps, then feeds stacked patch features to an image-wise CNN for global image-level classification.
Results
95% accuracy was achieved on the BACH four-class validation set, compared with previously reported 77% accuracy rates.
Takeaways & Limitations
The pipeline processes large-scale images using minimal hardware, with millisecond-order inference and relatively small pre-trained weights.
Takeaways & Limitations
Training the patch-wise network with image-level labels may disadvantage performance because not every patch represents the image category.
Abstract
from arXiv · showhide
This paper explores the problem of breast tissue classification of microscopy images. Based on the predominant cancer type the goal is to classify images into four categories of normal, benign, in situ carcinoma, and invasive carcinoma. Given a suitable training dataset, we utilize deep learning techniques to address the classification problem. Due to the large size of each image in the training dataset, we propose a patch-based technique which consists of two consecutive convolutional neural networks. The first "patch-wise" network acts as an auto-encoder that extracts the most salient features of image patches while the second "image-wise" network performs classification of the whole image. The first network is pre-trained and aimed at extracting local information while the second network obtains global information of an input image. We trained the networks using the ICIAR 2018 grand challenge on BreAst Cancer Histology (BACH) dataset. The proposed method yields 95 % accuracy on the validation set compared to previously reported 77 % accuracy rates in the literature. Our code is publicly available at https://github.com/ImagingLab/ICIAR2018
1 Introduction
The paper motivates automated four-class breast tissue classification from large, complex microscopy images, where pathologists face substantial diagnostic workload. It proposes a two-stage CNN pipeline that uses patch-wise local features and image-wise global information, achieving 95% validation accuracy on BACH.
- Breast cancer diagnosis is clinically important because early diagnosis can significantly increase treatment success.
- Microscopy images from biopsies are large and complex, increasing the workload of pathologists performing histopathological diagnosis.
- High-dimensional microscopy images motivate CNN-based classification, while large image sizes require patch-based processing because of hardware barriers.
- The proposed two-stage pipeline extracts salient local patch features before combining them with global information for image-level classification.
- 95% validation accuracy was achieved on the BACH dataset using the proposed model.
2 Related Works
Related work applies CNNs to breast cancer detection, classification, and segmentation, but studies address different class structures and tasks. The paper’s closest comparison is a four-class patch-based CNN followed by SVM and patch-probability fusion.
- The paper provides code and pre-trained weights through the ImagingLab/ICIAR2018 repository.
- Computer-aided diagnosis systems are described as fast and inexpensive alternatives to second-opinion methods, while deep learning has influenced medical image processing.
- Recent breast cancer CNN studies include two-class malignant-versus-benign classification, three-class classification, and segmentation.
- Araújo et al. are identified as the first team known to consider four-class breast tissue classification.
- Their method divided each image into twelve contiguous non-overlapping patches, used CNN or CNN+SVM patch probabilities, and fused them into image-level predictions.
3 Methods
The method processes high-resolution histology images through separate patch-wise and image-wise CNN stages. Overlapping patches support local feature extraction, while non-overlapping feature maps are stacked so the second network can learn image-level relationships.
- 3 Methods: The task is to classify 2048 × 1536 histology images into normal, benign, in situ carcinoma, and invasive carcinoma.
- 3 Methods: High-resolution inputs create memory constraints, while downsampling risks losing discriminative features such as nuclei information and density.
- Patch-Based Method with CNN: The patch-wise stage extracts fixed-size patches with a sliding k × k window and stride s before producing feature maps.
- Patch-Based Method with CNN: For image-wise training, s = 512 yields 12 non-overlapping patches whose feature maps are concatenated into a (12×C, 64, 64) 3D input.
- Network Architecture: The patch-wise CNN uses convolutional layers, batch normalization, ReLU activations, and stride-based dimensionality reduction to produce smaller feature maps.
- Network Architecture: The image-wise CNN uses convolutional downsampling, spatial averaging, fully connected layers, softmax classification, dropout, and early stopping.
4 Experiments and Results
Experiments use the balanced BACH histology dataset with augmentation and shared train/validation splits for both networks. The proposed method reaches 95.00% validation accuracy, with comparisons and channel-depth tests evaluating its design.
- Dataset: The BACH dataset contains 400 H&E-stained microscopy images, evenly distributed across four labels and annotated by two medical experts.
- Data Augmentation: Rotations, mirroring, and random color perturbations expand the patch-wise dataset to 16×35×400 samples.
- Training: The patch-wise model uses 80% of the dataset for training and 20% for cross-validation, with the same split used for the image-wise network.
- Results: 95.00% validation accuracy was obtained after an ensemble averaged eight rotation/flip input variations.
- Results: 93.75% validation accuracy and mean AUC 98.3 were achieved by the best non-ensemble model on 80 validation images.
- Comparison: The proposed patch-wise network outperformed previous methods in the comparison with Araújo et al. using patch-probability decision schemes.
- Ablation: A one-channel output configuration outperformed configurations with more channels in the reported validation experiments.
5 Conclusions and Future Work
The paper presents a two-stage CNN pipeline that combines patch-level and image-level processing to classify large microscopy images with limited hardware. It reports strong validation performance while acknowledging limitations from separate training and patch-label assumptions.
- Conclusions: The pipeline is designed to process large-scale images without the memory footprint required by end-to-end training.The authors identify minimal hardware use as the main contribution of the approach.
- Conclusions: The two-stage pipeline processes image patches with a patch-wise network and combines their feature maps with an image-wise network for global classification.The patch-wise network captures local features, while the image-wise network models relationships between neighboring patches and produces image-level scores.
- Results: 95% accuracy was achieved on the four-class validation set using the BACH dataset, compared with previously reported 77% accuracy.Inference took milliseconds, and the patch-wise and image-wise pretrained weights were 7.9MB and 1.6MB, respectively.
- Limitations: The patch-wise network was trained with image-level labels even though individual patches may not share the image’s category.The authors acknowledge that this labeling strategy may disadvantage performance.
- Future Work: End-to-end joint training was impractical in the authors’ experiments because it required a very large memory footprint.The authors identify investigation of this framework and its improvements as future work.