Source-linked AI summary
Breast Cancer Classification from Histopathological Images with Inception Recurrent Residual Convolutional Neural Network
Md Zahangir Alom, Chris Yakopcic, Tarek M. Taha, Vijayan K. Asari
TL;DR
Breast cancer histopathology classification requires evaluation across varied datasets, classes, magnifications, and image or patient-level settings. The paper applies IRRCNN to two public datasets and reports superior performance across the evaluated classification tasks and metrics, including accuracy, sensitivity, AUC, and ROC results.
Problem
The paper addresses breast cancer classification across binary and multi-class tasks using varied histopathology datasets and evaluation settings.
Method
The study applies the Inception Recurrent Residual Convolutional Neural Network to two benchmark datasets using image-level, patient-level, image-based, and patch-based analyses.
Results
IRRCNN achieved state-of-the-art testing accuracy on both datasets and reported 3.67% and 2.14% average-recognition improvements for image-level and patient-level multi-class classification.
Takeaways & Limitations
The reported results support IRRCNN as a high-performing approach for binary and multi-class breast cancer classification across the evaluated settings.
Abstract
from arXiv · showhide
The Deep Convolutional Neural Network (DCNN) is one of the most powerful and successful deep learning approaches. DCNNs have already provided superior performance in different modalities of medical imaging including breast cancer classification, segmentation, and detection. Breast cancer is one of the most common and dangerous cancers impacting women worldwide. In this paper, we have proposed a method for breast cancer classification with the Inception Recurrent Residual Convolutional Neural Network (IRRCNN) model. The IRRCNN is a powerful DCNN model that combines the strength of the Inception Network (Inception-v4), the Residual Network (ResNet), and the Recurrent Convolutional Neural Network (RCNN). The IRRCNN shows superior performance against equivalent Inception Networks, Residual Networks, and RCNNs for object recognition tasks. In this paper, the IRRCNN approach is applied for breast cancer classification on two publicly available datasets including BreakHis and Breast Cancer Classification Challenge 2015. The experimental results are compared against the existing machine learning and deep learning-based approaches with respect to image-based, patch-based, image-level, and patient-level classification. The IRRCNN model provides superior classification performance in terms of sensitivity, Area Under the Curve (AUC), the ROC curve, and global accuracy compared to existing approaches for both datasets.
Experimental setup
The study evaluates IRRCNN on two breast cancer datasets for binary and multi-class classification, using image- and patient-level analyses and alternative patch-selection strategies.
- Datasets and evaluation: IRRCNN was tested on BreakHis and the Breast Cancer Classification Challenge 2015 datasets for binary and multi-class classification.
- Patch processing: Images with whole-slide dimensions exceeding typical digital images were processed as multiple patches because some samples exceeded 2000 × 2000 pixels.
- Patch processing: Both random-crop and sequential non-overlapping patch-selection methods were considered.
Datasets
The experiments use two public histopathology datasets with different class structures, magnification settings, image resolutions, and evaluation inputs.
- BreakHis: BreakHis contains 7909 samples from 82 patients, divided into 2440 benign and 5429 malignant samples across 40×, 100×, 200×, and 400× magnifications.
- BreakHis: Figure 3 presents four benign tumor types in the first row and malignant tumors in the second row at 400× magnification.
- 2015 BC Classification Challenge: The 2015 Breast Cancer Classification Challenge dataset contains 249 high-resolution 2040×1536 H&E-stained images, with 229 used for training and the remainder for testing.
- 2015 BC Classification Challenge: Its four categories are normal tissue, benign tissue, in situ carcinoma, and invasive carcinoma.
- Evaluation inputs: For image-wise evaluation, inputs were resized to 128×128 pixels, augmented, or represented by 200 random patches combined with a Winner Take All method.
Data Augmentation
The study augments pathology images through geometric transformations and compares original, augmented, center-patch, and resized representations during training.
- Augmentation procedures: Augmentation used 40-degree rotation, width and height shifts, shearing, zooming, and horizontal and vertical flipping.
- Augmentation procedures: Because augmentation introduced noise in some image regions, the method also evaluated only the center patch of augmented samples.
- Optimization: Training used stochastic gradient descent with momentum 0.9 across three 50-epoch trials, reducing the learning rate by 10 after 50 epochs.
- Input representations: Figure 6 contrasts center patches and resized images from original samples with those from augmented samples.
Results and Discussion
The paper introduces automated binary and multi-class breast cancer classification with IRRCNN on two datasets and reports state-of-the-art testing accuracy for both.
- Overall results: IRRCNN was applied to automated binary and multi-class breast cancer classification on two different datasets.
- Overall results: The multi-class experiments considered four and eight classes.
- Overall results: The method achieved state-of-the-art testing accuracy for both datasets.
Results for BreakHis
The IRRCNN was evaluated on BreakHis using image-level and patient-level criteria for binary and eight-class classification across magnification factors. It achieved stronger performance than existing deep-learning and fusion-based approaches, with training accuracy highest at 100× magnification.
- Evaluation criteria: Image-level and patient-level criteria were used to evaluate binary and eight-class BreakHis classification across 40×, 100×, 200×, and 400× magnifications.Patient-level analysis aggregates correctly classified images by patient, while image-level analysis evaluates individual histopathological samples.
- Magnification analysis: 100× magnification produced the best training accuracy, with 200× a close second.The reported training and validation accuracy varied with sample magnification factor.
- Classification performance: The IRRCNN showed superior testing performance for both multi-class and binary classification compared with existing deep-learning methods.The multi-class results are reported in Table 3 and binary results in Table 4.
- Comparison with prior work: The proposed method outperformed the highest accuracy reported for the fusion methods evaluated in across all compared cases.The comparison used the highest accuracy reported in.
Results for Breast Cancer Classification Challenge 2015
On the 2015 Breast Cancer Classification Challenge dataset, IRRCNN performance was evaluated under resized, cropped, random-patch, and augmentation settings. Augmented resized samples performed best for training and validation, while augmented center patches and random patches achieved the strongest testing results.
- Training and validation: Resized and augmented samples achieved the highest training and validation accuracy among the evaluated input preparations.The compared preparations included resized and augmented, center-patch cropped and augmented, random patches, resized, and center-patch cropped samples.
- Patch-wise classification: Augmented center patches achieved the highest patch-wise testing accuracy: 97.51% for binary and 97.11% for multi-class classification.Random patches showed similar performance, whereas single center patches reached 88.7% and 88.12%, respectively.
- Image-wise classification: Single resized images produced the lowest testing accuracy in the image-wise experiments.The passage contrasts this setting with random patches using WTA and augmented resized samples.
Analysis and Comparison Against State-of-the-Art
The paper compares IRRCNN with earlier machine-learning and deep-learning methods on BreakHis and the 2015 challenge dataset. It reports higher recognition accuracy across the evaluated classification settings, alongside computational-time measurements per sample.
- BreakHis comparison: On BreakHis, IRRCNN achieved 97.95±1.07% image-level and 97.65±1.20% patient-level accuracy for benign versus malignant classification.These results improved average performance over the highest accuracies reported in by 1.05% and 0.55%, respectively.
- BreakHis comparison: For BreakHis eight-class classification, IRRCNN reached 97.57±0.89% image-level and 96.84±1.13% patient-level accuracy.The reported improvements over the latest performance in were 3.67% and 2.14%, respectively.
- 2015 challenge comparison: On the 2015 challenge dataset, IRRCNN achieved 99.05% binary and 98.59% multi-class testing accuracy with resized and augmented samples.The paper also reports 100% testing performance for an experiment using random patches.
- Scope of evaluation: The study evaluates IRRCNN across image-level, patient-level, image-based, and patch-based analyses on two benchmark datasets.The experiments vary magnification, resizing, cropping, random patches, and augmentation criteria.