# Development of automatic organ segmentation based on positron-emission tomography analysis system using Swin UNETR in breast cancer patients in Korea

## TL;DR

This study introduces a deep learning system to automatically segment organs in PET scans of breast cancer patients, improving the accuracy of SUV measurements for better prognosis.

## Contribution

A novel deep learning-based organ segmentation system using Swin UNETR is developed to standardize SUV evaluation in breast cancer patients.

## Key findings

- The method achieved an auto-segmentation accuracy of 0.9311 across key organs in 10 patients.
- Automated SUV analysis showed improved reliability compared to traditional single-ROI methods with differences of 0.19 and 0.16 for maximum and mean SUV values.

## Abstract

The standardized uptake value (SUV) is a key quantitative index in nuclear medicine imaging; however, variations in region‐of‐interest (ROI) determination exist across institutions. This study aims to standardize SUV evaluation by introducing a deep learning‐based quantitative analysis method that enhances diagnostic and prognostic accuracy.

We used the Swin UNETR model to automatically segment key organs (breast, liver, spleen, and bone marrow) critical for breast cancer prognosis. Tumor segmentation was performed iteratively based on predefined SUV thresholds, and prognostic information was extracted from the liver, spleen, and bone marrow (reticuloendothelial system). The artificial intelligence training process employed 3 datasets: a test dataset (40 patients), a validation dataset (10 patients), and an independent test dataset (10 patients). To validate our approach, we compared the SUV values obtained using our method with those produced by commercial software.

In a dataset of 10 patients, our method achieved an auto‐segmentation accuracy of 0.9311 for all target organs. Comparison of maximum SUV and mean SUV values from our automated segmentation with those from traditional single‐ROI methods revealed differences of 0.19 and 0.16, respectively, demonstrating improved reliability and accuracy in whole‐organ SUV analysis.

This study successfully standardized SUV calculation in nuclear medicine imaging through deep learning‐based automated organ segmentation and SUV analysis, significantly enhancing accuracy in predicting breast cancer prognosis.