Two-Stage Convolutional Neural Network for Breast Cancer Histology Image Classification (1803.04054v2)

Abstract: 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

• The paper introduces a two-stage CNN architecture that significantly improves classification accuracy, achieving 95% on the BACH dataset.
• It employs a patch-wise network for local feature extraction followed by an image-wise network for holistic analysis.
• The method addresses computational inefficiencies with large histology images, offering a promising tool for automated breast cancer diagnostics.

Two-Stage Convolutional Neural Network for Breast Cancer Histology Image Classification

The paper "Two-Stage Convolutional Neural Network for Breast Cancer Histology Image Classification" presents a novel approach to the classification of breast cancer histological images. The primary objective is to effectively categorize images into four distinct types: normal, benign, in situ carcinoma, and invasive carcinoma, using deep learning techniques. The focus on these categories provides critical insights for potential applications in medical diagnostics and automating pathological analysis.

Methodology

The authors introduce a two-stage convolutional neural network (CNN) architecture that addresses the challenge of large image sizes in the provided dataset. This approach is a significant departure from traditional methods, which have reported lower accuracy rates. The process begins with patch extraction from the whole slide images, addressing the computational inefficiencies that arise from the sheer scale of the data.

1. Patch-wise Network: The initial stage operates as an auto-encoder, facilitating the extraction of salient features from image patches. This network is pre-trained to emphasize local information, ensuring that the subsequent analysis is driven by relevant and significant visual features.
2. Image-wise Network: Building on the output of the patch-wise network, the second stage executes holistic image classification. By integrating the features identified in the first stage, this network achieves a comprehensive understanding of global image contexts, crucial for accurate categorization into one of the four cancer types.

This two-tiered CNN leverages the ICIAR 2018 BACH dataset, a respected benchmark in breast cancer histology image classification. The substantial improvements reported in accuracy—reaching 95% on the validation set compared to the earlier cited performance of 77%—demonstrate the enhanced efficacy of this method.

Results and Implications

The reported results underscore the robustness of the proposed method, with a marked improvement in classification accuracy over existing approaches. The careful structuring of the two-stage CNN, with its pre-trained patch-wise component, illustrates the importance of feature extraction at different image scales. These findings have meaningful implications for breast cancer diagnostics in that they posit a reliable, automated tool for pathologists, potentially accelerating diagnostic processes and broadening access to high-quality healthcare solutions.

Furthermore, the network's strong performance on the BACH dataset suggests potential applicability across similar medical imaging challenges. The architecture could provide a template for tackling other histological classification tasks, fostering advancements within medical image processing disciplines.

Future Directions

The paper concludes with potential avenues for further research, inspired by the promising outcomes of the initial paper. Future developments could explore the refining of patch extraction processes or integrating novel training paradigms to enhance model generalizability. Moreover, expanding the dataset or employing transfer learning could provide additional benchmarks to validate the utility of this approach across diverse medical imaging scenarios.

In conclusion, the proposed two-stage convolutional neural network provides a significant advancement in the field of medical image classification, particularly concerning breast cancer histology. By delivering a method characterized by increased accuracy and efficiency, the researchers contribute valuable insights into the deployment of AI in healthcare diagnostics, presenting a framework that holds promise for further innovations and practical applications in the domain.