Source-linked AI summary
Convolutional neural network classification of cancer cytopathology images: taking breast cancer as an example
MingXuan Xiao, Yufeng Li, Xu Yan, Min Gao, Weimin Wang
TL;DR
Breast cancer histopathology diagnosis requires specialized expertise and time. This paper applies fixed Inception-V3 features, quadtree image partitioning, and block-probability fusion to classify images. Its reported conclusion is that segmentation improves recognition accuracy, with four-way splitting outperforming 16-way splitting.
Problem
Breast histopathology diagnosis relies on expert visual examination, which requires substantial specialized expertise.
Source: §1 · p. 1Method
The method classifies breast histopathology images using fixed Inception-V3 features, quadtree partitioning, and sum, product, or maximum fusion of block predictions.
Source: §3 · pp. 2–3Results
Image segmentation significantly enhances recognition accuracy, four-way splitting performs better than 16-way splitting, and fusion algorithms have relatively minor impact on segmented-image recognition rates.
Source: §5 · p. 4Takeaways & Limitations
The reported findings support image segmentation as an effective component for improving breast-image recognition accuracy in this study.
Source: §5 · p. 4
Abstract
from arXiv · showhide
Breast cancer is a relatively common cancer among gynecological cancers. Its diagnosis often relies on the pathology of cells in the lesion. The pathological diagnosis of breast cancer not only requires professionals and time, but also sometimes involves subjective judgment. To address the challenges of dependence on pathologists expertise and the time-consuming nature of achieving accurate breast pathological image classification, this paper introduces an approach utilizing convolutional neural networks (CNNs) for the rapid categorization of pathological images, aiming to enhance the efficiency of breast pathological image detection. And the approach enables the rapid and automatic classification of pathological images into benign and malignant groups. The methodology involves utilizing a convolutional neural network (CNN) model leveraging the Inceptionv3 architecture and transfer learning algorithm for extracting features from pathological images. Utilizing a neural network with fully connected layers and employing the SoftMax function for image classification. Additionally, the concept of image partitioning is introduced to handle high-resolution images. To achieve the ultimate classification outcome, the classification probabilities of each image block are aggregated using three algorithms: summation, product, and maximum. Experimental validation was conducted on the BreaKHis public dataset, resulting in accuracy rates surpassing 0.92 across all four magnification coefficients (40X, 100X, 200X, and 400X). It demonstrates that the proposed method effectively enhances the accuracy in classifying pathological images of breast cancer.
1 INTRODUCTION
Breast cancer diagnosis relies heavily on histopathological examination and pathologists’ expertise. The paper presents automated image categorization as a way to address the time and expertise involved in accurate diagnosis.
1 INTRODUCTION
- Histopathological examination is the predominant diagnostic method, and most diagnoses rely on pathologists’ visual examination of specimens.The text notes that this work requires substantial specialized expertise.Source: §1 · p. 1
- Expert diagnostic concordance averages approximately 75%, underscoring the role of pathologists’ experience in breast cancer detection.Source: §1 · p. 1
- Automated computational categorization of histopathological images is presented as a means to expedite diagnosis and mitigate errors.Source: §1 · p. 1
2 RELATED WORK
Prior breast cancer image-classification work includes manual-feature methods and deep-learning approaches. The paper uses the BreaKHis dataset, which contains H&E-stained images from four magnification levels.
2 RELATED WORK
- Traditional methods combine manually extracted features with classifiers; reported accuracy ranged from 70% to 100% for Belsare [2] and 80% to 85% for Spanhol [17][18].Source: §2 · p. 1
- Deep-learning studies include CNN classification and magnification-agnostic approaches, while this study adds image segmentation for high-resolution images.Prior reported results include 77.8% overall accuracy for Zhang [19], maximum accuracies of 90% and 86.3% for Spanhol [16], and 84.3% for Bayramoglu.Source: §2 · p. 2
2.1 Data Source
- BreaKHis contains 7,909 H&E-stained images from 82 patients, with 2,480 benign and 5,429 malignant samples across 40X, 100X, 200X, and 400X magnifications.Source: §2.1 · p. 2, Table 1 · p. 2
3 METHODOLOGY
The method uses fixed Inception-V3 features and a fully connected classifier, with quadtree image partitioning to expand training data and fusion rules to combine block predictions.
3 METHODOLOGY
- Images are resized to 299×299, processed with fixed ImageNet-pretrained Inception-V3 weights, and classified from 2,048-dimensional features by two fully connected layers and SoftMax.The original model’s last two fully connected layers are excluded; the replacement layers have 512 and 2 nodes.Source: §3 · p. 2
- For training, cross-entropy between SoftMax outputs and image labels is minimized by gradient descent on the fully connected network.Source: §3 · pp. 2–3
- Quadtree partitioning divides images into 4 or 16 blocks at levels two and three, multiplying training data by 4 or 16 while assigning each block its source image’s label.The paper uses this segmentation-based augmentation to address overfitting risk when datasets have few samples.Source: §3 · p. 3
- The method combines predictions from image blocks using sum, product, or maximum fusion to obtain a classification result.The three rules are tested as alternatives for consolidating block-level classifications.Source: §3 · p. 3
4 EXPERIMENTAL
Experiments evaluated image- and patient-level accuracy on BreaKHis, comparing image segmentation strategies and the proposed approach with prior methods. Segmentation improved accuracy over non-split images, while the authors note that further verification is constrained by dataset size.
4.1 Experimental Design
- The experimental design divides the dataset into training, validation, and testing sets, reported as 75%, 15%, and 15%, respectively.Image-level accuracy counts correctly diagnosed images over all images; patient-level accuracy averages accuracy for each patient.Source: §4.1 · pp. 3–4
4.2 Experimental Results
- Image-level accuracy rose 2.0%–4.4% and patient-level accuracy 0.9%–4.9% with segmentation versus non-split images.Quarter-split segmentation had higher accuracy than 16-way splitting in 79.2% of cases; the authors attribute this to greater retained detail.Source: Table 2 · p. 4, Table 3 · p. 4, §4.2 · p. 4
- The proposed approach improved image-recognition accuracy by 9.9%–12.2% versus manual-feature methods.Reported gains were 2.3%–5.1% versus rotation-and-flipping augmentation, 5%–11.5% versus sliding-window augmentation, 8.2%–10.9% versus deep-feature fusion, and 10.2%–12% versus a multi-task CNN.Source: §4.2 · p. 4
- The authors say validation of the reported accuracy is constrained by the limited dataset and needs substantially more data.They also conclude that CNN feature extraction outperforms manual feature extraction in classification performance.Source: §4.2 · p. 4
5 CONCLUSION
The conclusion reports improved breast-image recognition with segmentation and better performance for 4-way than 16-way splits. It also identifies finer tumor categorization and algorithm improvements as directions for further work.
5 CONCLUSION
- Image segmentation significantly enhances breast-image recognition accuracy, with 4-way split outperforming 16-way split.The conclusion also reports that different fusion algorithms have relatively minor effects on recognition rates for segmented images.Source: §5 · p. 4
- Deep-learning feature extraction is reported to outperform visually based manual feature extraction in the comparisons discussed.Source: §5 · p. 4
- Further work is proposed on feature extraction, segmentation, and data augmentation, alongside finer categorization beyond benign and malignant classes.Source: §5 · p. 4