Title: Leveraging Visual Mamba and a 3D Segment Anything Model for the Detection of Placenta Accreta Spectrum

Placenta Accreta Spectrum (PAS) constitutes a rare yet perilous obstetric condition. Timely and precise identification of PAS is essential for safeguarding maternal well-being. Currently, diagnosis typically depends on the judgment of seasoned clinicians who evaluate Magnetic Resonance Imaging (MRI) scans alongside a patient’s cesarean delivery history. However, many district-level hospitals face significant hurdles in achieving accurate diagnoses due to insufficient expertise and limited resources.

To mitigate these disparities, we have compiled the inaugural MRI-based dataset dedicated to PAS, featuring comprehensive annotations for both classification and fine-grained segmentation. Furthermore, we posit that isolating lesion regions from uterine MRI scans can substantially bolster diagnostic accuracy.

In pursuit of automated PAS diagnosis, we introduce 3DSAMba, an innovative framework designed for robust feature learning and effective lesion segmentation. Our approach begins with the development of a 3D Segment Anything Model (SAM), which is augmented with medical-specific insights via a streamlined adapter mechanism. Additionally, the framework employs a Multi-Level Aggregation Mamba (MLAM) to consolidate feature maps across various hierarchical levels, alongside a Fusion State Space Model (FSSM) that integrates multi-scale features derived from both the encoder and decoder stages.

By applying the generated segmentation masks to the initial MRI images through element-wise multiplication, the method effectively extracts lesion zones, thereby facilitating more precise PAS detection. Rigorous experimental results confirm that this framework markedly enhances diagnostic outcomes for PAS. To support ongoing research in this field, we have made the dataset and source code publicly available at https://github.com/Drchip61/PASD.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC