A Novel Approach to Classify Breast Cancer using Transfer Learning

Breast cancer is the second leading cause of death among women. However, accurate early diagnosis can successfully increase the survival rate. This study proposes a pre-trained technique to classify mammogram images into cancerous or not cancerous. A new model has been proposed based on the pre-trained convolutional neural network architectures—InceptionV3, ResNet50, DenseNet-121, EfficientNetB0, and EfficientNetB3 are known for classifying images accurately. Image augmentation with further pre-processing techniques and hyperparameter tuning was employed to improve overall model accuracy and prevent overfitting. The MIAS dataset is used that contains images of three categories normal, benign, and malignant. Experimental results show that EfficientNetB3 provides a superior classification accuracy of 100% on training, 99.47% on testing, and the right prediction is 99.05%. The findings of this study are expected to be useful in cancer diagnosis research and possibly assist radiologists in clinical decision-making.