A Convolutional Neural Network Approach for Deep Learning-based Breast Cancer Detection

Abstract

Early and accurate breast cancer (B) diagnosis is crucial for improving patient outcomes and survival rates, which currently range from 30% to 50%. Deep learning has emerged as a powerful tool in healthcare, particularly for analyzing large volumes of medical images like X-rays, MRIs, and CT scans. This study introduces a novel deep learning model, BCNN, designed to detect and classify breast cancers into Three distinct categories: malignant lobular carcinoma, malignant mucinous carcinoma, and malignant papillary carcinoma. The BCNN model was developed and compared against two fine-tuned, pre-trained models which is VGG16, MobileNet initially trained on the ImageNet database. Breast cancer MRI images, sourced from a Kaggle dataset, were used for training and evaluation. To enhance the dataset's size and diversity, a Generative Adversarial Network technique was employed for data augmentation. The dataset included images at four different magnifications (40X, 100X, 200X, and 400X), along with a combined dataset. Each model, including the proposed BCNN, was evaluated across all five datasets, resulting in a total of 30 experiments. Performance was assessed using F1-score, recall, precision, and accuracy metrics. The experimental results demonstrated the effectiveness of the proposed BCNN model, achieving a classification F1-score accuracy of 99.38%. The fine-tuned pre-trained models also performed well, with the following F1-score accuracies: VGG16 produce 97.67%), and MobileNet Produce 97.38%). The study concluded that data augmentation, preprocessing, and balancing significantly improved the performance of both the BCNN model and the fine-tuned pre-trained models. Notably, the highest accuracies were observed when using the 400X magnification images, likely due to their superior resolution.