Breast cancer is the uncontrolled cell growth in the breast and is the second most common cancer among women in the United States. There are around 240,000 new cases per year reported among women in the United States, and there are around 42,000 deaths per year related to breast cancer in the United States. The number of new cases and deaths have been increasing due to the increasing population size.

Doctors save lives by segmenting out regions of interest (like tumors) in ultrasound images. However, segmentation can be a long and tedious process if one wants to segment by hand.

Segment Anything Model (SAM) is an AI model from Meta AI that can segment out any object within any image. It is easy to manipulate and can be incorporated into other systems. It accepts multiple input prompts such as points, boxes, and masks, which is useful in this project.

Despite its many strengths in segmenting regular images, SAM is not that accurate when segmenting ultrasound images. Ultrasound images contain a lot of noise that make it difficult for SAM to segment accurately. In addition, SAM is trained on a large diverse dataset including images not specific to ultrasound.

The dice and cross-entropy loss decreased significantly over the model training process. ClickSAM achieved a mean Intersection over Union (IoU) of 0.916. MedSAM achieved a mean IoU of 0.863. Segmentation Click Train achieved a mean IoU of 0.707. Additional prompts (i.e., clicks) provided during the second training improved the quality of the segmentation drastically.

Click prompts: precise coordinates, intricate definition of desired areas, each segment can be assigned both positive and negative prompts

Bounding boxes: only rectangular regions, each segment can be assigned only one bounding box (no positive or negative boxes), axis-aligned nature