A Step Closer Toward Precision Medicine by Leveraging a Deep Learning Model to Detect Knee Osteoarthritis: Our Experience with a Large Data Set of 6,571 Patients

Presentation Format: Oral presentation.

Purpose or Learning Objective: Knee osteoarthritis is one of the most common, painful joint diseases, without a definitive treatment to reverse it. Most automated models that predict the Kellgren-Lawrence (KL) grade of osteoarthritic knee radiographs were developed from data sets sourced outside India. Therefore, an automatic system robust enough to predict KL grade on Indian knee radiographs was developed.

Methods or Background: A total of 6,571 cases, 4,447 Osteoarthritis Initiative (OAI), 1,043 Indian Institute 1 and 1,081 Indian Institute 2, knee radiographs, anteroposterior view, of patients aged 14 to 83 years, were used in the study. Apart from sides, joint segmentation became a crucial first step due to irrelevant features like femur and fibula in the radiographs. Overall, 1,000 random radiographs from OAI and 589 random images from Indian Institute 2 were annotated into left and right knee joints by expert radiologists. They were split into train (1,112 images), validation (159 images), and test (318 images) sets to train a Mask R-CNN model to segment knee joints automatically.

The Mask R-CNN model was used to extract knee joints from all the images and was graded using the KL system by expert radiologists.

These segmented knee joints were further used to train a regression model based on DenseNet-121 to automatically predict the KL grade from an image. The regression model was trained on the OAI data set and fine-tuned using the Indian data sets.

Results or Findings: The regression model correctly classified 66.96% of the cases in the private test set where 32.66% of the misclassifications were among neighboring KL grades. Mean absolute error of 0.2097, precision of 0.78, and recall of 0.67 were obtained.

Conclusion: Regression considers the inherent ordering between the KL grades and significantly decreases misclassification in nonadjacent classes. Automated and accurate assignment of KL grades to knee radiographs can help mitigate the effects of human subjectivity in assessing radiographs, reduce radiologist workload, and improve reporting times. Also, one single model is unlikely to work across different cohorts, but we can fine-tune it to the concerned cohort.

No conflict of interest has been declared by the author(s).

Publication History

Article published online:
03 June 2021