Automatic magnetic resonance prostate segmentation by deep learning with holistically nested networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Prostate segmentation in MRI is crucial but challenging due to anatomical variability and image artifacts.
• Traditional methods struggle with accuracy because of noise and similar tissue intensities near the prostate boundary.
• Deep learning offers potential for improved segmentation accuracy.

Purpose of the Study:

• To develop and evaluate advanced deep learning methods for accurate automatic prostate segmentation in MRI.
• To compare a novel holistic deep learning approach with a patch-based method.
• To establish a new state-of-the-art in MRI prostate segmentation.

Main Methods:

• Investigated patch-based convolutional networks for prostate contour refinement.
• Proposed an end-to-end segmentation method integrating Holistically Nested Edge Detection (HNED) with Fully Convolutional Networks (FCNs).
• Evaluated the enhanced Holistically Nested Network (HNN) model on 250 patient MRI scans using fivefold cross-validation.

Main Results:

• The enhanced HNN model achieved a Dice Similarity Coefficient (DSC) of [Formula: see text] and a Jaccard Index (IoU) of [Formula: see text].
• The holistic HNN model significantly outperformed a patch-based AlexNet model, showing a 9% improvement in DSC and 13% in IoU.
• The method demonstrated state-of-the-art performance compared to existing MRI prostate segmentation techniques.

Conclusions:

• The proposed enhanced HNN model provides highly accurate automatic prostate segmentation in MRI.
• Holistic deep learning methods, particularly the enhanced HNN, offer significant advantages over patch-based approaches.
• This research advances the field of medical image analysis for prostate cancer diagnosis and treatment planning.