Breast cancer classification using attention-based transfer learning on histopathological images

Breast cancer is one of the most prevalent and deadly cancers affecting women worldwide. Early and accurate diagnosis plays a crucial role in improving survival rates and guiding effective treatment strategies. This research aims to tackle the challenges of accurate breast cancer subtype classification by applying and optimizing advanced deep learning models, ultimately providing an effective solution that improves diagnostic reliability through the use of attention mechanisms and comprehensive data augmentation. We propose and comprehensively evaluate several deep learning-based approaches for the multiclass classification of breast cancer subtypes using histopathological images. The experiments were conducted on the BreakHis dataset, which originally contained 7,909 histopathological images categorized into eight subtypes of benign and malignant breast tumors. To enhance the diversity and robustness of the training data, data augmentation techniques were employed, increasing the total number of training images to 19,960. For all experiments in this study, we used an 80/20 split of the dataset, assigning 80% of the images for training and reserving the remaining 20% for testing. This approach ensured that every model was evaluated under the same conditions. Several advanced Convolutional Neural Network (CNN) models were explored, including ResNet152V2, MobileNetV2, and versions of these models that included attention modules like Squeeze-and-Excitation (SE) and Convolutional Block Attention Module (CBAM). Each model was trained both with and without data augmentation. After training, their performance was measured using a range of metrics accuracy, precision, recall, F1-score, and confusion matrices broken down for each tumor type. The results were clear: adding attention mechanisms and using data augmentation almost always led to better accuracy, regardless of the model or class being tested. Out of all the models, CBAM-MobileNetV2 stood out, achieving the highest test accuracy of %98.35. It also performed better than the other models in terms of how well it generalized and handled the different tumor subtypes. This work shows how deep learning and especially CNNs with attention mechanisms can be powerful tools for accurately and automatically classifying breast cancer subtypes from pathology images. The findings also highlight how crucial it is to use effective data augmentation and thoughtful data splitting strategies, since these steps help address issues with class imbalance and make the models more reliable, which is especially important if these tools are to be used in real-world medical settings.