Automatic segmentation of prostate MRI based on 3D pyramid pooling Unet

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Automatic segmentation of prostate magnetic resonance (MR) images is vital for diagnosing and evaluating prostate diseases, including cancer.
• Convolutional neural networks are the current standard for prostate segmentation, but limitations in spatial context modeling hinder performance.
• Existing methods struggle with the complex tissue structures in MR images, necessitating improved segmentation techniques.

Purpose of the Study:

• To develop a novel 3D pyramid pool Unet to enhance prostate MR image segmentation.
• To address limitations in spatial context modeling and improve segmentation accuracy in complex MR images.
• To achieve more precise segmentation for better prostate cancer diagnosis and management.

Main Methods:

• Proposed a 3D pyramid pool Unet incorporating pyramid pooling in skip connections (SC) and deep supervision (DS).
• Modified SC to merge decoder layers with same-scale and smaller-scale encoder feature maps, combining low-level details and high-level semantics.
• Utilized pyramid pooling for multifaceted feature extraction and DS for hierarchical representation learning to boost accuracy.

Main Results:

• Experiments on 78 patient 3D prostate MR images showed high correlation with expert manual segmentation.
• Achieved an average relative volume difference of 2.32% and a Dice similarity coefficient of 91.03% for prostate volume segmentation.
• Demonstrated superior performance compared to existing segmentation methods.

Conclusions:

• The proposed 3D pyramid pool Unet achieves highly accurate prostate MR image segmentation.
• The method's results are quantitatively consistent with expert manual segmentation.
• This approach offers a promising advancement for clinical applications in prostate cancer diagnosis and evaluation.