Is the fairness of segmentation foundation models sufficient for clinical cardiology applications (RF_TH_165)

Thursday, January 30, 2025

11:20 AM – 11:30 AM East Coast USA Time

Location: Blue Room PreFunction

Presenting Author(s)

QL

Qing Li, MSc

Shanghai, China
Human Phenome Institute, Fudan University, China (People's Republic)

Primary Author(s)

QL

Qing Li, MSc

Shanghai, China
Human Phenome Institute, Fudan University, China (People's Republic)

Co-Author(s)

YZ

Yizhe Zhang, PhD

Nan Jing, China
School of Computer Science, Nanjing University of Science and Technology, China (People's Republic)
AX

Aijia Xie, MSc

Shanghai, China
Human Phenome Institute, Fudan University, China (People's Republic)
LS

Longyu Sun, MSc

Shanghai, China
Human Phenome Institute, Fudan University, China (People's Republic)
MS

Mengting Sun, MSc

Shanghai, China
Human Phenome Institute, Fudan University, China (People's Republic)
SW

Shuo Wang, PhD

Shanghai, China
Digital Medical Research Center, School of Basic Medical Sciences, Fudan University, China (People's Republic)
Chengyan Wang, PhD

Associate Professor
Fudan University, China (People's Republic)

Background:

Foundation models for segmentation have shown tremendous potential in medical image segmentation as well as the clinical applications in cardiac. However, there is limit in study on the assessment of foundation models in cardiac image segmentation which is crucial to help develop models that can achieve fair segmentation effects across different demographic groups to enhance diagnostic accuracy in clinical settings. In this paper, we assess the fairness of the foundation model SAM and Medical SAM (MedSAM) in cardiac image segmentation.

Methods:

A total of 701 healthy volunteers (289 males and 412 females) between June 2023 and February 2024 were included in the study(Figure 1.a). The average age of the subjects was 36 ± 12 years with the average body mass index (BMI) of 23.35 ± 3.46. Data were acquired using a 3T scanner (MAGNETOM Vida, Siemens Healthineers), equipped with dedicated multi-channel cardiac coils. 8 heart regions(ventricle and atrium) and 9 vessels were segmented using foundation models (Figure 1.b). The DSC (Dice Similarity Coefficient) was