Automated measurement and grading of knee cartilage thickness: a deep learning-based approach

JiangRong Guo, Pengfei Yan, Yong Qin, MeiNa Liu, Yingkai Ma, Jing Li, Ren Wang, Hao Luo, Songcen LvFrontiers in Medicine 2024 · DOI

Abstract

By selecting the best knee cartilage segmentation network, we built a model with a stronger generalization ability to automatically segment, measure, and grade cartilage thickness. This model can assist surgeons in more accurately and efficiently diagnosing changes in patients' cartilage thickness.

Introduction

Knee Osteoarthritis (KOA) is a leading cause of disability worldwide. While traditional X-ray imaging is the standard diagnostic tool, it often fails to detect damage until significant cartilage loss has occurred, leading to bone-on-bone contact. MRI offers the potential for much earlier detection by visualizing cartilage thinning years before it becomes apparent on X-rays. However, the manual measurement of cartilage thickness from MRI scans is labor-intensive, subjective, and impractical for high-volume clinical settings. This study presents a robust deep learning pipeline designed to automate the segmentation, measurement, and grading of knee cartilage.

Methods

The researchers developed a multi-stage automated system. The first stage involved comparing four deep learning architectures for cartilage segmentation: UNet, UNet++, ResUNet, and TransUNet. The TransUNet model was chosen for its ability to capture long-distance dependencies through a Transformer-based encoder, which is particularly effective for the thin, continuous structure of knee cartilage.

Following segmentation, the system processes the masks to identify femoral and tibial cartilage. To standardize the results, the knee is divided into 12 anatomical regions based on the WORMS (Whole Organ Magnetic Resonance Imaging Score) scale, allowing for granular thickness calculations across the joint.

Anatomical regions for cartilage thickness measurement.

The model was trained on a diverse "mixed" dataset of 991 MRI cases from four sources: a local Hospital dataset, the Osteoarthritis Initiative (OAI), fastMRI (NYU), and SKI10. Additionally, a healthy baseline was established by screening over 7,000 cases to find 291 "normal" subjects, providing a regional standard for comparison.

Results

TransUNet outperformed the other architectures, achieving an overall Dice Similarity Coefficient (DSC) of 0.813 and an Intersection over Union (IoU) of 0.692. A key finding was that training on a mixed dataset significantly improved the model's generalization across different MRI hardware and protocols compared to models trained on a single source.

Segmentation performance across models and datasets.

The automated measurement system showed a Mean Absolute Percentage Error (MAPE) of 10.10% relative to manual expert measurements. Most importantly, the model achieved an accuracy of 89.99% in automatically grading the severity of cartilage wear (Grades 1–4). The study also confirmed biological benchmarks, noting that cartilage is generally thicker in men and in right knees.

Discussion and Limitations

This pipeline provides a scalable tool for early KOA diagnosis by establishing a standardized "yardstick" for cartilage health. However, the authors note several limitations. The current "normal" baseline is specific to a population in Northeast China and may require adjustment for other demographic groups. Furthermore, the model does not yet incorporate patient-specific covariates such as BMI or activity levels. Positioned as an auxiliary tool, this system is intended to enhance the efficiency and accuracy of surgeons in clinical practice rather than replace expert judgment. Future work will focus on expanding the dataset to create a more globally representative standard for knee health.

Summary updated

Abstract

By selecting the best knee cartilage segmentation network, we built a model with a stronger generalization ability to automatically segment, measure, and grade cartilage thickness. This model can assist surgeons in more accurately and efficiently diagnosing changes in patients' cartilage thickness.
BibTeX
@inproceedings{guo2024automated,
  author = {JiangRong Guo and Pengfei Yan and Yong Qin and MeiNa Liu and Yingkai Ma and Jing Li and Ren Wang and Hao Luo and Songcen Lv},
  title = {{Automated measurement and grading of knee cartilage thickness: a deep learning-based approach}},
  booktitle = {Frontiers in Medicine},
  year = {2024},
  doi = {10.3389/fmed.2024.1337993},
}
Figure