MixUNETR: A U-shaped network based on W-MSA and depth-wise convolution with channel and spatial interactions for...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate prostate cancer diagnosis and staging rely on precise Magnetic Resonance Imaging (MRI) segmentation of the peripheral zone (PZ) and transition zone (TZ).
• Existing segmentation methods struggle with ambiguous boundaries, shape variations, and texture complexities between PZ and TZ, limiting diagnostic accuracy and AI-driven analysis.
• Inadequate modeling capabilities and limited receptive fields in current approaches hinder effective feature extraction for prostate MRI segmentation.

Purpose of the Study:

• To develop an enhanced segmentation method for prostate MRI that accurately delineates the PZ and TZ.
• To address the limitations of existing methods in handling complex boundaries and feature extraction.
• To improve the accuracy and robustness of artificial intelligence-driven prostate cancer analysis through superior MRI segmentation.

Main Methods:

• Proposed Enhanced MixFormer integrating window-based multi-head self-attention (W-MSA) and depth-wise convolution with parallel design and cross-branch bidirectional interaction.
• Introduced MixUNETR, utilizing multiple Enhanced MixFormers as an encoder for comprehensive feature extraction from both PZ and TZ in prostate MRI.
• Augmented receptive field and modeling capability to enhance extraction of global and local features for improved segmentation.

Main Results:

• MixUNETR demonstrated superior accuracy and robustness in segmenting prostate MRI compared to state-of-the-art methods.
• The method effectively addressed challenges in delineating boundaries between PZ and TZ, reducing mis-segmentation.
• Consistent performance was observed across Prostate158, ProstateX public datasets, and a private dataset.

Conclusions:

• MixUNETR offers a significant advancement in prostate MRI segmentation, enhancing diagnostic capabilities.
• The proposed model effectively overcomes limitations of previous methods, improving the delineation of critical prostate zones.
• This work provides a robust solution for accurate AI-driven prostate cancer analysis, with code available for reproducibility.