CNN-based fully automatic wrist cartilage volume quantification in MR images: A comparative analysis between...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Orthopedics

Background:

• Accurate measurement of wrist cartilage volume is crucial for diagnosing and monitoring joint diseases.
• Manual segmentation of wrist cartilage in MR images is time-consuming and prone to inter-observer variability.
• Developing automated methods for cartilage segmentation can improve efficiency and consistency in clinical practice.

Purpose of the Study:

• To evaluate the performance of various deep learning architectures, including U-Net variants and Mask R-CNN, for automatic wrist cartilage segmentation in MR images.
• To compare the accuracy and reproducibility of these automated methods against a previously developed patch-based CNN and manual segmentation.
• To assess the influence of image parameters on the segmentation reproducibility.

Main Methods:

• Four manually optimized U-Net variants, nnU-Net, and Mask R-CNN frameworks were employed for wrist cartilage segmentation.
• Segmentation quality was evaluated by comparison with manual segmentation using metrics such as Dice Similarity Coefficient (DSC) and volume error.
• A cross-validation approach was used on a dataset of 33 3D VIBE MR images, with a focus on healthy volunteers.

Main Results:

• U-Net-based networks demonstrated superior segmentation homogeneity and quality compared to the patch-based CNN.
• The U-Net with attention layers (U-Net_AL) achieved the highest median 3D DSC (0.817) and the best correlation with ground truth (Pearson's r = 0.765).
• U-Net_AL exhibited lower mean volume error (17%) and higher reproducibility than manual segmentation, though image resolution significantly impacted results.

Conclusions:

• Convolutional neural networks (CNNs) with attention layers (U-Net_AL) provide the most effective method for wrist cartilage segmentation.
• The trained U-Net model can be fine-tuned for specific patient groups to facilitate clinical implementation.
• Independent validation using non-MRI methods is recommended to assess the overall error in cartilage volume measurement.