MonoUNet: A Robust Tiny Neural Network for Automated Knee Cartilage Segmentation on Point-of-care Ultrasound Devices

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Orthopedics

Background:

• Point-of-care ultrasound (POCUS) is increasingly used for musculoskeletal assessments.
• Accurate segmentation of knee cartilage is crucial for diagnosing and monitoring knee osteoarthritis (OA).
• Existing deep learning models for cartilage segmentation can be computationally intensive and less robust to variations in ultrasound image quality.

Purpose of the Study:

• To develop a compact and robust deep learning model for automated knee cartilage segmentation specifically for POCUS devices.
• To evaluate the performance and reliability of the proposed model against manual segmentation and existing lightweight models.

Main Methods:

• A novel, compact deep learning model named MonoUNet was developed, featuring a reduced U-Net backbone and a trainable monogenic block for multi-scale feature extraction.
• A gating mechanism was incorporated to enhance robustness to variations in ultrasound image appearance.
• The model was evaluated on a multi-site, multi-device knee cartilage ultrasound dataset, using Dice score, mean average surface distance (MASD), Bland-Altman analysis, and intraclass correlation coefficient (ICC2,k).

Main Results:

• MonoUNet achieved superior segmentation performance compared to existing lightweight models, with average Dice scores of 92.62%–94.82% and MASD values of 0.133–0.254 mm.
• The model demonstrated significant reductions in parameters (10×–700×) and computational cost (14×–2000×).
• Excellent agreement and reliability were observed between MonoUNet and manual cartilage measurements (ICC2,k=0.96 for thickness, ICC2,k=0.99 for echo intensity).

Conclusions:

• The integration of trainable local phase features enhances the robustness of compact neural networks for knee cartilage segmentation.
• MonoUNet offers a scalable solution for ultrasound-based assessment and monitoring of knee OA using POCUS devices.
• The model's efficiency and accuracy support its potential for widespread clinical adoption in POCUS settings.