[Paper Review] Breast Cancer Classification from Histopathological Images with Inception Recurrent Residual Convolutional Neural Network

Md Zahangir Alom, Chris Yakopcic+2 more|arXiv (Cornell University)|Nov 10, 2018

AI in cancer detection15 references4 citations

TL;DR

This paper proposes an Inception Recurrent Residual Convolutional Neural Network (IRRCNN) for breast cancer classification from histopathological images, integrating Inception-v4, ResNet, and recurrent CNN components to enhance feature learning and classification accuracy. The model achieves state-of-the-art performance on BreakHis and 2015 Breast Cancer Classification Challenge datasets, outperforming existing methods in sensitivity, AUC, ROC area, and global accuracy.

ABSTRACT

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.

Motivation & Objective

• To improve breast cancer classification accuracy from histopathological images using deep learning.
• To address limitations in existing models by combining the strengths of Inception, Residual, and Recurrent CNN architectures.
• To evaluate the proposed IRRCNN model on both image-level and patient-level classification tasks.
• To compare the IRRCNN performance against state-of-the-art machine learning and deep learning methods across multiple evaluation metrics.
• To demonstrate superior performance in sensitivity, AUC, and global accuracy on public datasets.

Proposed method

• The IRRCNN model integrates the Inception-v4 architecture for multi-scale feature extraction and efficient computation.
• It incorporates residual connections from ResNet to mitigate vanishing gradient problems and improve training depth.
• Recurrent convolutional layers are introduced to model spatial dependencies across feature maps, enhancing contextual understanding.
• The model is trained end-to-end on histopathological whole-slide images and patch-level representations.
• Training uses stochastic gradient descent with cross-entropy loss, and data augmentation techniques are applied to improve generalization.
• The framework supports both image-level and patient-level classification by aggregating predictions across multiple patches.

Experimental results

Research questions

• RQ1 Can a hybrid deep learning architecture combining Inception, residual, and recurrent components improve breast cancer classification accuracy on histopathological images?
• RQ2 How does the IRRCNN model compare to standalone Inception, ResNet, and RCNN models in terms of sensitivity and AUC?
• RQ3 Does the integration of recurrent connections enhance feature representation in histopathological image classification?
• RQ4 What is the performance of IRRCNN on both image-level and patient-level classification tasks across benchmark datasets?
• RQ5 Can the proposed model generalize effectively to diverse histopathological image datasets like BreakHis and the 2015 Breast Cancer Classification Challenge?

Key findings

• The IRRCNN model achieved the highest global accuracy on the BreakHis dataset compared to existing deep learning and machine learning methods.
• On the 2015 Breast Cancer Classification Challenge dataset, the IRRCNN demonstrated superior sensitivity and AUC values, outperforming baseline models.
• The model achieved an AUC of 0.98 on the BreakHis dataset, indicating excellent discrimination between benign and malignant cases.
• The IRRCNN showed consistent improvements in ROC curve performance across both datasets, confirming robustness.
• The integration of recurrent layers contributed to better feature representation, especially in capturing long-range spatial dependencies in tissue patterns.
• The model achieved state-of-the-art performance in both image-level and patient-level classification, highlighting its clinical relevance.