MambaX-Net leverages prior scans for improved longitudinal prostate MRI segmentation

MambaX-Net, a semi-supervised dual-scan 3D segmentation model that uses prior MRI and segmentation mask, outperformed U-Net and Transformer baselines for longitudinal prostate zone segmentation on active surveillance MRI, even with limited and noisy labels.

• Architecture combines a Mamba-enhanced Cross-Attention Module and a Shape Extractor Module to encode the previous time point's MRI and segmentation mask for refined zone delineation.
• Uses semi-supervised self-training with pseudo-labels generated from a pre-trained nnU-Net, enabling learning without expert annotations.
• Evaluated on a longitudinal active surveillance dataset and significantly outperformed state-of-the-art U-Net and Transformer-based models.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.