BMA-Net: A 3D bidirectional multi-scale feature aggregation network for prostate region segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate prostate MRI segmentation is vital for diagnosis.
• Transformers improve global feature representation but are computationally intensive, limiting them to single slices.
• Existing methods struggle to integrate intra-slice and inter-slice multi-scale information effectively.

Purpose of the Study:

• To develop a 3D network that effectively utilizes both intra-slice and inter-slice multi-scale information for precise prostate MRI segmentation.
• To address the limitations of single-slice processing in Transformer-based segmentation models.

Main Methods:

• Proposed BMA-Net, a 3D bidirectional multi-scale feature aggregation network.
• Employed a frequency-based global feature filtering branch for intra- and inter-slice information.
• Integrated a spatial attention branch and a CNN branch for local features, with multi-scale feature fusion.

Main Results:

• Achieved Dice coefficients of 88.35% (central gland) and 76.86% (peripheral zone) on a public dataset.
• Obtained Dice coefficients of 85.85% (central gland) and 74.50% (peripheral zone) on an in-house dataset.
• Demonstrated superior segmentation performance compared to state-of-the-art methods.

Conclusions:

• BMA-Net effectively leverages multi-scale information for enhanced prostate segmentation accuracy.
• The proposed network achieves superior performance over existing state-of-the-art methods in prostate MRI segmentation.