A deep convolutional neural network for the classification of imbalanced breast cancer dataset

Abstract

The primary procedures for breast cancer diagnosis involve the assessment of histopathological slide images by skilled patholo-gists. This procedure is prone to human subjectivity and can lead to diagnostic errors with adverse implications for patient health and welfare. Artificial intelligence-based models have yielded promising results in other medical tasks and offer tools for potentially addressing the shortcomings of traditional medical image analysis. The BreakHis breast cancer dataset suffers from insufficient data for the minority class with an imbalance ratio >0.40, which poses challenges for deep learning models. To avoid performance degradation, researchers have explored a variety of data augmentation schemes to generate adequate samples for analysis. This study designed a Deep Convolutional Neural Network (DCGAN) with specific generator and discriminator architectures to mitigate model instability and generate high-quality synthetic data for the minority class. The balanced dataset was passed to the fine-tuned ResNet50 model for breast tumor detection. The study produced high accuracy in diagnosing benign/malignancy at 40X magnification, outperforming the state-of-art. The results demonstrated that deep learning methods can potentially to support effective screening in clinical practice.

Keywords

• Deep convolutional neural network
• Breast cancer histopathological images
• Data augmentation
• Deep convolutional generative adversarial network
• Frechet inception distance evaluation criterion
• BreakHis dataset