Abstract:
The study utilizes deep learning, primarily transfer learning models, to enhance breast
cancer diagnosis using ultrasound images. The Breast Ultrasound images (BUI)
collection included 780 images classified as benign (487), malignant (210), and normal
(133). The dataset was augmented to 5760 images to improve model training. Four pretrained transfer learning models-VGG16, ResNet50, Xception, and a DenseNet201
were tested for their ability to classify breast cancer images. Before feeding image to
the model this study employed several image pre-processing techniques like image
resizing, gaussian filter, normalization, gamma correction to enhance the image quality.
Xception achieved the best accuracy of 99%, outperforming the other models in this
testing. The accuracy values for VGG16 were 96%, ResNet50 at 90%, and
DenseNet201 at 93%. Despite its success, the study had limitations. The original
dataset's size and variety are limited, which may have an influence on the model's
applicability in real-world circumstances.
