Non-overlapping Patch-Based Pre-trained CNN for Breast Cancer Classification

Breast cancer (BC) significantly impacts women's mortality rates and requires early detection to improve survival chances and enable appropriate treatment. Thus, a computer-aided system with high performance can speed up this process. A convolutional neural network (CNN) is considered sensitive to insufficient, noisy data. It cannot achieve high performance, however, restricted access to high-quality medical data, stemming from stringent confidentiality and privacy issues, is a considerable obstacle to the successful training of deep learning models. The current study aims to develop a remarkable, influential model for BC classification whilst considering modern pre-trained models ResNet50, AlexNet, InceptionV3 and VGG16 for extracting feature vectors of images. Furthermore, to improve the quality of images, some preprocessing steps are performed, and augmentation is done to solve the imbalanced problem of the dataset. Non-overlapping patching is applied during preprocessing, where each image is divided into smaller, distinct patches. This strategy allows the model to focus on localized histopathological features, reducing background noise and improving feature extraction. Unlike traditional methods that process whole images, this approach enhances the ability to capture subtle cancer-related patterns, contributing to better classification performance. A support vector machine and a neural network are implemented to classify the extracted features of BC histopathological images. Results obtained from the current paper are critical. Under BreaKHis dataset, 32 models are trained using nonoverlap patching of four different image resolutions (40×, 100×, 200× and 400×). After training the models, ResNet50 with the NN classifier had the most accurate results, with an accuracy of 98.02% and sensitivity of 98.01% at 200× magnification, which outperformed all models. The ResNet50 with SVM had 97.3% accuracy and 97.3% sensitivity, which is more than 7% higher than the baseline model. The results indicate that integrating localised patch-based analysis into deep learning frameworks might enhance diagnostic precision in histopathological image evaluation, perhaps aiding pathologists in the early and more dependable diagnosis of breast cancer. This presents the effectiveness of preprocessing and augmentation techniques.