3D Segment Anything Model With Visual Mamba for Diagnosing Placenta Accreta Spectrum

Area of Science:

• Obstetrics and Gynecology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Placenta Accreta Spectrum (PAS) is a rare but dangerous obstetric condition requiring early diagnosis.
• Current PAS diagnosis relies on expert analysis of cesarean history and MRI, which is challenging in resource-limited hospitals.
• Accurate PAS diagnosis is crucial for maternal health outcomes.

Purpose of the Study:

• To develop an automated system for accurate Placenta Accreta Spectrum diagnosis using MRI.
• To create the first MRI-based PAS dataset with segmentation and classification annotations.
• To enhance lesion segmentation in uterine MRI for improved PAS detection.

Main Methods:

• Proposed 3DSAMba, a novel feature learning framework for effective lesion segmentation in MRI.
• Developed a 3D Segment Anything Model (SAM) with medical domain adaptation via an adapter mechanism.
• Introduced Multi-Level Aggregation Mamba (MLAM) and Fusion State Space Model (FSSM) for advanced feature aggregation and fusion.

Main Results:

• The 3DSAMba framework significantly improved diagnostic performance for Placenta Accreta Spectrum.
• Lesion areas were effectively isolated using segmentation masks applied to MRI images.
• The developed dataset and source code are publicly available to advance PAS research.

Conclusions:

• The 3DSAMba framework offers a promising approach for automated and accurate PAS diagnosis.
• AI-powered lesion segmentation in MRI can overcome resource limitations in diagnosing PAS.
• This work facilitates further research and clinical application in obstetric imaging analysis.