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

A research team from multiple institutions has developed MonoUNet, a breakthrough tiny neural network specifically designed for automated knee cartilage segmentation on point-of-care ultrasound (POCUS) devices. The model incorporates three key innovations: an aggressively reduced U-Net backbone with asymmetric decoder, a trainable monogenic block that extracts multi-scale local phase features, and a gated feature injection mechanism that integrates these features into encoder stages. This architecture dramatically improves robustness across different ultrasound devices and acquisition settings while maintaining exceptional accuracy.

MonoUNet was rigorously evaluated on a multi-site, multi-device knee cartilage ultrasound dataset acquired using cart-based, portable, and handheld POCUS devices. The results are impressive: average Dice scores ranging from 92.62% to 94.82% and mean average surface distance values between 0.133 mm and 0.254 mm. Most remarkably, the model achieves these results while being 10x-700x smaller in parameters and 14x-2000x more computationally efficient than existing lightweight segmentation models. The cartilage measurements showed excellent reliability with intraclass correlation coefficients of 0.96-0.99 when compared to manual expert annotations.

The practical implications are significant for scalable osteoarthritis monitoring. By enabling accurate, automated cartilage segmentation on portable ultrasound devices, MonoUNet could transform how knee osteoarthritis is assessed in primary care settings, sports medicine clinics, and remote locations. The model's tiny footprint makes it suitable for deployment on resource-constrained devices, potentially democratizing access to quantitative cartilage assessment tools that were previously limited to specialized imaging centers with expensive MRI equipment.