Probability Map-Guided Network for 3D Volumetric Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• 3D medical images offer rich anatomical context but suffer from anisotropy and intensity inhomogeneities.
• Anisotropy can cause blurring and distortion, while inhomogeneities obscure lesion boundaries and introduce noise.
• Accurate segmentation of 3D medical images is crucial for diagnosis and treatment planning.

Purpose of the Study:

• To propose a probability map-guided network (3D-PMGNet) for accurate 3D volumetric medical image segmentation.
• To address challenges posed by anisotropy and intensity inhomogeneities in 3D medical imaging.
• To enhance the reliability and detail of segmentation results in complex medical scans.

Main Methods:

• Developed a 3D-PMGNet utilizing probability maps from intermediate features as supervisory signals.
• Introduced a novel probability map reconstruction method combining dynamic thresholding and local adaptive smoothing.
• Incorporated a learnable channel-wise temperature coefficient and dynamic prompt encoding for feature fusion and adjustment.

Main Results:

• The proposed 3D-PMGNet demonstrated superior performance compared to state-of-the-art methods on four diverse datasets.
• The probability map guidance effectively mitigated issues related to image anisotropy and intensity variations.
• Experimental results validated the enhanced accuracy and robustness of the proposed segmentation approach.

Conclusions:

• 3D-PMGNet offers a robust solution for 3D volumetric medical image segmentation, outperforming existing methods.
• The novel probability map reconstruction and feature fusion techniques are key to the method's success.
• The publicly released source code facilitates further research and application in medical image analysis.