According to the World Health Organization, breast cancer is one of the forms of cancer with the highest prevalence worldwide. It is important that the diagnostic process is improved as much as possible. 

The goal of this project was to determine the best way to train image segmentation models on breast ultrasound datasets, and to produce fully trained segmentation models that perform well on real-world data. 

Combining several small ultrasound datasets into a single large dataset was an important part of the project, ensuring the models can be trained with sufficient data to attain the desired levels of performance. 

 Both traditional segmentation models (U-Net – based on a ResNet CNN) and state-of-the-art models (SegFormer – a segmentation transformer) were trained for this project. Extensive performance optimizations were performed for the main models via hyperparameter optimization (Optuna). 

 This work was done within a larger project at the University of Wisconsin, under the guidance of Dr. Jeff Baggett, and the models and artifacts generated here may be integrated in various ways within the parent project. The output from the segmentation models could be used, within ensemble methods, to provide input for other models to perform further predictions. Or the predicted segmentation masks could be used directly in an app, providing visual cues to radiologists as they are performing ultrasound scans in a real-life scenario.