# Practical application of SAM for breast nodules segmentation

## TL;DR

This paper explores how the Segment Anything Model (SAM) can be used to segment breast nodules in medical images, aiming to improve breast cancer detection.

## Contribution

The study evaluates SAM's performance on breast nodule segmentation using custom data and proposes practical approaches for better results.

## Key findings

- Using MedSAM initial weights improves segmentation accuracy for breast nodules.
- Fixed prompt boxes help achieve better and more practical segmentation outcomes.
- The study confirms the feasibility of SAM for breast nodule segmentation in real-world scenarios.

## Abstract

Breast cancer is one of the most common and deadly diseases that threaten women’s health worldwide, and early and accurate breast nodule segmentation is of great significance for the early detection, diagnosis and treatment of breast cancer. However, due to the limitation of medical annotated data, the training segmentation models for medical images is still challenging. The Segment Anything Model (SAM) is a foundational model that interactively segments target objects. Although significant achievements have been made in natural images, there are still challenges in the application in medical images. In this paper, the effect of SAM on breast nodule segmentation was studied from three aspects: initial weight, organ (breast) mask and prompt box, so as to explore the feasibility of breast nodule segmentation. Through a series of experiments on the data collected in this paper, it is found that the use of MedSAM initial weights and the use of single individual fixed prompt boxes can obtain better segmentation results, and can take into account practical application problems.