Automated Grading of Osteoarthritis on the KL Scale from Knee Radiographs
Osteoarthritis (OA) is the most prevalent form of arthritis commonly affecting the knees, spine, hips, and hands. Knee osteoarthritis is most observed in people who are above the age of 45, suffer from obesity, or lead a sedentary lifestyle. In India, the prevalence of osteoarthritis lies between 22% and 39%*. These figures are expected to rise considering the trends of longevity and obesity. Since damage caused by the onset of osteoarthritis is irreversible, the current focus of doctors is on symptom control in the early stages of osteoarthritis.
This blog article describes the work that is published in detail at:
https://arxiv.org/abs/2004.08572
Kondal Sudeep, Viraj Kulkarni, Ashrika Gaikwad, Amit Kharat, and Aniruddha Pant. "Automatic Grading of Knee Osteoarthritis on the Kellgren-Lawrence Scale from Radiographs Using Convolutional Neural Networks." arXiv preprint arXiv:2004.08572 (2020).
Despite the availability of advanced medical imaging techniques such as Magnetic Resonance Imaging (MRI) and Computed Tomography (CT), the knee X-ray remains the most widely used modality for the diagnosis of knee OA. The severity of osteoarthritis is graded on the Kellgren-Lawrence scale: 0 for no arthritis, 1 for stage 1 arthritis, and continuing till 4 for severe arthritis.
Human evaluation of X-rays inevitably introduces some subjectivity in diagnosis. Computer-aided diagnosis can mitigate this subjectivity by automating the process and leaving the final decision up to the radiologists. DeepTek’s deep learning models do just this by automatically detecting the severity of knee osteoarthritis cases from X-rays, thereby reducing the workload on radiologists and helping them provide timely diagnoses.
We began our work by developing a pre-processing model that automatically crops and labels the 4447 knee X-rays we had obtained from the Osteoarthritis Initiative’s open-source repository. The Osteoarthritis Initiative is a research study sponsored by the National Institutes of Health. Our model helped crop out additional features present in the X-rays like the femur and fibula, which were irrelevant to our grading model. It also separated images containing both the knees and marked them as either left or right. Our in-house radiologists had manually annotated the training images this model was built on.
Once we had our segmented knee joints, we built both classification and regression grading models. We trained the classification model using cross-entropy as the loss function. In contrast, we trained the regression model using mean squared error as the loss function to predict the KL grade as a real number. The model’s output was rounded off to the closest integer ranging from 0 to 4 to obtain the final predicted grade.
In the last leg of development, we fine-tuned these models on X-rays obtained from a well-known private Indian hospital. Our evaluation of the models found that only 68% of all misclassifications were amongst neighbouring grades for the classification model, while 87% of all misclassifications were amongst neighbouring grades for the regression model. These observations support treating the KL grade as an ordinal variable rather than a nominal one.
Since different stages of osteoarthritis require different interventions, it is important to identify the grade of osteoarthritis afflicting a patient. By automatically assigning KL grades to knee X-ray images with performance matching that of expert radiologists, DeepTek’s AI models mitigate the effects of human subjectivity in assessing X-rays, reduce radiologist work burden, and improve report turnaround time.
This automated KL grading system is accessible to diagnostic centres via DeepTek’s cloud-based platform. Using our model, doctors can quickly identify the severity of knee osteoarthritis affecting a patient and then prescribe the most effective treatment based on this.
