Automatic Meniscus Segmentation Using Adversarial Learning-Based Segmentation Network with Object-Aware Map in Knee...

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Accurate meniscus segmentation in knee MR images is vital for quantitative analysis and 3D model reconstruction, particularly for meniscus allograft transplantation.
• Existing segmentation methods may struggle with accuracy and efficiency, necessitating advanced deep learning approaches.
• The patient's normal meniscus geometry is used to create a 3D model for allograft transplantation.

Purpose of the Study:

• To propose a novel two-stage deep convolutional neural network (DCNN) for automated and accurate meniscus segmentation in knee MR images.
• To enhance the precision of meniscus segmentation for improved 3D reconstruction and surgical planning in meniscus allograft transplantation.
• To address challenges like class imbalance and segmentation inaccuracies (under/over-segmentation) using advanced deep learning techniques.

Main Methods:

• A two-stage DCNN approach combining a 2D U-Net for meniscus localization and a conditional generative adversarial network (cGAN) for segmentation.
• Stage 1: 2D U-Net segments knee MR images into six classes (including bone and cartilage) at 512x512 resolution to localize medial and lateral menisci.
• Stage 2: Adversarial learning with a U-Net-based generator and DCNN-based discriminator, utilizing an object-aware map on localized 64x64 regions for refined segmentation.

Main Results:

• Achieved average Dice similarity coefficients of 85.18% for the medial meniscus and 84.33% for the lateral meniscus.
• Demonstrated significant improvements over methods without adversarial learning (10.79%p and 7.78%p higher) and without object-aware maps (1.14%p and 1.12%p higher).
• The multi-class localization effectively prevented class imbalance, while adversarial learning with object-aware maps reduced under- and over-segmentation.

Conclusions:

• The proposed two-stage DCNN with adversarial learning and object-aware maps provides accurate and robust meniscus segmentation in knee MR images.
• This automated method facilitates precise shape analysis of the meniscus, essential for creating patient-specific 3D models for allograft transplantation.
• The approach offers a significant advancement in medical image analysis for orthopedic surgery planning and patient-specific treatments.