Enhanced Breast Cancer Detection using ResNet50V2-based Convolutional Neural Networks
Pratham Kaushik, Sunila Choudhary · 2024
Breast cancer is one of the leading causes of death among the female population globally; hence, there is a need for diagnostic tools with high accuracy and reliability. Conventionally used diagnostic techniques, which include mammography, ultrasound, and histopathological examination, are proven methods; however, they have disadvantages such as interpretative variability and require highly skilled pathologists. Recently developed deep learning algorithms, especially Convolutional Neural Networks, have shown huge potential for the automation of medical image analysis. This research study proposes a new model for breast cancer detection based on the ResNet50V2 architecture, which was pre-trained for image classification tasks on the ImageNet dataset but is now being used to classify histopathological images from the BreakHis dataset. The proposed model uses data augmentation techniques like random brightness adjustment, flipping, and rotation for optimum generalization before extraction of features using the ResNet50V2 base. It is then fine-tuned with fully connected layers and dropout to prevent overfitting. The performance of the model for the 25 epochs ran to a ROC-AUC of 0.84413, an accuracy of 0.79634, and a loss of 0.47527. While very promising, this showed a high number of false negatives for the model, very indicative that it requires further calibration and testing on larger and more diverse datasets. This paper, therefore, represents the potential of ResNet50V2-based models to improve aid to pathologists for the automated detection of breast cancer, consequently opening avenues toward better prognosis in patients.