Abstract:
Breast cancer is one of the most common types of malignancies seen in women,
hence initiating a huge amount of health, psychological, social, and economic
consequences in the lives of patients, their families, and also in healthcare systems.
Early Detection is Key to Better Outcome Medical imaging, especially ultrasound
is a very broad based and non-invasive diagnostic tool. However, its accuracy
commonly relies on radiologist expertise, which may vary. Machine learning and
deep learning approaches provide a promising way to not only improve the
reliability and efficiency of diagnosis but also in resource constrained
environments. This work supports the combination of machine learning with
ultrasound imaging for proper cancer classification and diagnosis of breast
ultrasound imaging. A large dataset of Ultrasound images was expanded so that it
balances all the classes that reaches around ten thousand samples with the help of
augmentation techniques like random resized crop, Gaussian noise to ensure
balanced representation. Several deep learning architectures were trained and
evaluated such as ResNet50, Hybrid (CoatNet), ViT Base, Swin Tiny, EfficientNet
B3, DeiT Small, DeiT Base Distilled, MaxViT Tiny, MaxViT Base, and RepViT M1.
Here MaxViT Tiny showed the best validation accuracy of 90.72% with test
accuracy result of 85.38%. On the other hand, the accuracy rate of the DeiT Base
Distilled model was 87.99% on validation data sets and 87.93% on test data sets.
But DeiT Base Distilled model shows more consistent performance thus it’s been
chosen as the best model. Based on these results, a mobile application with deep
learning power is proposed. Using a Flutter interface and a FastAPI backend
deployed on the Hugging Face Spaces, the system will attempt to provide faster,
more accurate, and accessible breast cancer diagnostics especially in low-resource
settings.
