Abstract
Achilles tendon rupture is a debilitating injury, typically treated with surgical repair and long-term rehabilitation. The recovery, however, is protracted and often incomplete. Diagnosis and healing progress assessment are largely based on ultrasound and magnetic resonance imaging. In this paper, we propose an automatic method based on deep learning for analysis of Achilles tendon condition and estimation of its healing progress on ultrasound images. We develop custom convolutional neural networks for classification and regression on healing score and feature extraction. Our models are trained and validated on an acquired dataset of over 250,000 sagittal and over 450,000 axial ultrasound slices. The obtained estimates show high correlation with the assessment of expert radiologists, with respect to all key parameters describing healing progress. We also observe that parameters associated with intratendinous healing processes are better modeled with sagittal slices. We prove that ultrasound imaging is quantitatively useful for clinical assessment of Achilles tendon healing process and should be viewed as complementary to magnetic resonance imaging.
Introduction
Achilles tendon ruptures are severe injuries with notoriously long recovery periods. Currently, clinical monitoring relies heavily on qualitative assessments of Ultrasound (US) or Magnetic Resonance Imaging (MRI) scans. While MRI offers high resolution, ultrasound is more accessible and cost-effective. However, US images are often "noisy," containing speckle noise and artifacts that make manual interpretation subjective and automated analysis challenging. This research introduces a deep learning framework to quantitatively score tendon healing using standard 2D ultrasound scans, bridging the gap between qualitative observation and objective measurement.
Datasets and Methodology
The study utilized a massive longitudinal dataset comprising over 700,000 2D ultrasound slices extracted from 3D exams of 49 patients and 18 healthy volunteers. These slices were categorized into sagittal (longitudinal) and axial (cross-sectional) views.
The researchers employed a two-stage approach:
- Binary Classification: Training CNNs (AlexNet, ResNet50, and Inception-v3) to distinguish between healthy and injured tissue.
- Healing Progress Estimation: Modifying these architectures for regression to predict scores for six clinical parameters—such as structural changes, tendon thickening, and edema—on a 7-point scale.
Key Insights
- The Importance of Perspective: The study found that different clinical markers are better captured in specific planes. Sagittal slices were superior for internal structural changes and thickening, while axial slices were more effective for assessing edge sharpness and tissue edema.
- Contextual Information: Interestingly, cropping images to focus solely on the tendon (ROI segmentation) reduced model accuracy. The surrounding tissue provides essential context for the network to gauge injury severity.
- Supervised vs. Semi-Supervised: While semi-supervised methods like PCA have worked for MRI, they failed for ultrasound due to inherent noise. This necessitated an end-to-end supervised regression approach.
Results and Performance
The Inception-v3 model achieved a high classification accuracy of 91.6% for sagittal slices. In the regression task, the models produced Mean Absolute Errors (MAE) between 0.67 and 1.08 on the 7-point scale, demonstrating strong alignment with expert radiologist scores.
The automated estimates showed significant correlation with clinical ground truth, particularly for parameters like tendon thickening and structural changes. This suggests that AI can provide a reliable, objective second opinion for clinicians.
Conclusion
This work demonstrates that ultrasound imaging, when combined with deep learning, is a powerful quantitative tool for monitoring Achilles tendon healing. By providing objective scores that correlate with expert assessments, this framework could standardize rehabilitation monitoring and improve clinical outcomes. Future efforts will focus on generalizing these models across different ultrasound hardware and conducting clinical trials to evaluate AI-assisted diagnosis.
Summary updated