Estimating Achilles Tendon Healing Progress with Convolutional Neural Networks

Norbert Kapinski, Jakub Zielinski, Bartosz A. Borucki, Tomasz Trzcinski, Beata Ciszkowska-Lyson, Krzysztof S. NowinskiMICCAI 2018 · DOI · OpenReview

Abstract

Quantitative assessment of a treatment progress in the Achilles tendon healing process - one of the most common musculoskeletal disorder in modern medical practice - is typically a long and complex process: multiple MRI protocols need to be acquired and analysed by radiology experts. In this paper, we propose to significantly reduce the complexity of this assessment using a novel method based on a pre-trained convolutional neural network. We first train our neural network on over 500,000 2D axial cross-sections from over 3000 3D MRI studies to classify MRI images as belonging to a healthy or injured class, depending on the patient's condition. We then take the outputs of modified pre-trained network and apply linear regression on the PCA-reduced space of the features to assess treatment progress. Our method allows to reduce up to 5-fold the amount of data needed to be registered during the MRI scan without any information loss. Furthermore, we are able to predict the healing process phase with equal accuracy to human experts in 3 out of 6 main criteria. Finally, contrary to the current approaches to regeneration assessment that rely on radiologist subjective opinion, our method allows to objectively compare different treatments methods which can lead to improved diagnostics and patient's recovery.

Introduction

Achilles tendon rupture is a prevalent musculoskeletal injury with a significant risk of re-rupture, estimated between 20% and 40%. Monitoring the healing process is critical for clinical management, yet current methods are inefficient. The gold standard involves acquiring up to 10 distinct MRI protocols, which are then subjectively evaluated by radiologists across multiple criteria. This process is not only time-consuming and expensive but also prone to inter-observer variability. This research introduces an automated, objective framework to quantify healing progress using deep learning, aiming to maintain diagnostic accuracy while significantly reducing data requirements.

Methodology

The proposed system utilizes a multi-stage pipeline designed to extract clinically relevant information from high-dimensional MRI data.

  1. Binary Pre-training: Standard CNN architectures (AlexNet, ResNet-18, and GoogleNet) were first trained on a large-scale dataset of 2D axial MRI slices to distinguish between "healthy" and "injured" tendons.
  2. Feature Extraction: The AlexNet model was selected as a feature extractor, with activations pulled from the first fully connected layer (fc6).
  3. Dimensionality Reduction: Principal Component Analysis (PCA) was applied to the high-dimensional feature vectors. Analysis revealed that the first principal component (PC1) captured over 50% of the variance associated with the healing continuum.
  4. The $H$ Metric: A novel metric, $H$, was defined as the truncated mean of PC1 values across all slices in a 3D MRI study, serving as a global indicator of tendon condition.
  5. Regression: Linear regression mapped the $H$ metric to 5-point clinical scores provided by expert radiologists.
The proposed pipeline: MRI slices are processed through a CNN feature extractor (FE), followed by fully connected layers (FC) and PCA to derive the H metric.

Data and Experimental Setup

The study utilized a specialized dataset acquired from a GE Signa HDxt 1.5T scanner:

  • Healthy Cohort: 270 3D MRI scans from 27 volunteers.
  • Injured Cohort: 2,772 3D MRI scans from 60 patients, tracked longitudinally over 10 time points (pre-surgery to 52 weeks post-operation).
  • Total Data: The training set comprised 234,502 healthy and 277,208 injured 2D slices.
  • Ground Truth: Radiologists scored each scan on a 1–5 scale for six parameters: Structural changes (SCT), Tendon thickening (TT), Sharpness of edges (STE), Tendon edema (TE), Tendon uniformity (TU), and Tissue edema (TisE).

Results and Discussion

The experimental results demonstrate the efficacy of using mid-level neural representations for clinical assessment. The pre-trained AlexNet achieved a binary classification accuracy of 99.19%. More importantly, the derived $H$ metric showed strong clinical relevance.

The model matched human-level precision, predicting healing phases with a Mean Square Error (MSE) of less than 0.5 points on the 5-point scale. Specifically, the $H$ metric exhibited high Pearson correlation with three key clinical criteria: Sharpness of tendon edges (STE), Tendon edema (TE), and Tissue edema (TisE).

Correlation between predicted values and ground truth for three clinical criteria: Sharpness of edges (STE), Tendon edema (TE), and Tissue edema (TisE).

A major finding of this work is the potential for significant efficiency gains. By identifying redundancies between MRI protocols, the researchers demonstrated that high-quality assessments could be achieved using only 2 protocols (PD and T2*GRE) instead of the standard 10, representing a 5-fold reduction in data acquisition time.

Conclusion and Future Work

While the current 2D axial approach is highly effective for many parameters, it faces limitations in assessing "Tendon Uniformity" (TU), which typically requires sagittal views. Future iterations will explore incorporating 3D context and additional principal components to improve predictions for structural changes and tendon thickening. Nonetheless, this objective, automated approach provides a scalable alternative to subjective radiologist evaluation, potentially standardizing the comparison of different treatment modalities.

Summary updated

Abstract

Quantitative assessment of a treatment progress in the Achilles tendon healing process - one of the most common musculoskeletal disorders in modern medical practice - is typically a long and complex process: multiple MRI protocols need to be acquired and analysed by radiology experts for proper assessment. In this paper, we propose to significantly reduce the complexity of this process by using a novel method based on a pre-trained convolutional neural network. We first train our neural network on over 500 000 2D axial cross-sections from over 3 000 3D MRI studies to classify MRI images as belonging to a healthy or injured class, depending on the patient’s condition. We then take the outputs of a modified pre-trained network and apply linear regression on the PCA-reduced space of the features to assess treatment progress. Our method allows to reduce up to 5-fold the amount of data needed to be registered during the MRI scan without any information loss. Furthermore, we are able to predict the healing process phase with equal accuracy to human experts in 3 out of 6 main criteria. Finally, contrary to the current approaches to healing assessment that rely on radiologist subjective opinion, our method allows to objectively compare different treatments methods which can lead to faster patient’s recovery.
BibTeX
@inproceedings{kapinski2018estimating,
  author = {Norbert Kapinski and Jakub Zielinski and Bartosz A. Borucki and Tomasz Trzcinski and Beata Ciszkowska-Lyson and Krzysztof S. Nowinski},
  title = {{Estimating Achilles Tendon Healing Progress with Convolutional Neural Networks}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2018},
  doi = {10.1007/978-3-030-00934-2\_105},
}
Figure