E1539. Automated Convolutional Neural Network Grading of Knee Osteoarthritis on the Kellgren-Lawrence Scale from Radiographs in Indian Population
  1. Amit Kharat; DeepTek Imaging Private Limited
  2. Sudeep Kondal; DeepTek Imaging Private Limited
  3. Priyam Choudhury; DeepTek Imaging Private Limited
  4. Ashrika Gaikwad; DeepTek Imaging Private Limited
  5. Viraj Kulkarni; DeepTek Imaging Private Limited
  6. Aniruddha Pant; DeepTek Imaging Private Limited
Knee osteoarthritis is one of the most common, painful joint diseases without a definitive reversal treatment. The severity of the disease is graded on Kellgren-Lawrence (KL) scale (1), 0 being healthy knee and 1-4 being graded on increasing severity of affliction. However, the grading suffers from radiologists subjectivity while reading films (2). Most automated models developed to date are from datasets outside India, predicting KL grade of the osteoarthritic knee radiographs, which didn’t perform well on Indian knee X-Rays. Therefore, both regression and classification novel models were designed, which performed better on Indian radiographs.

Materials and Methods:
4447 Osteoarthritis Initiative (OAI) and 1043 Indian institute (Target) knee radiographs, AP view, of patients in the age range 45 -79 years were used in the study. Apart from differentiating into left and right sides, joint segmentation became a crucial first step due to irrelevant features like femur and fibula in the radiographs. 1000 random radiographs from OAI were annotated into left and right knee joints by expert radiologists. They were then split into the train (700 images), validation (100 images) and test (200 images) sets, to train a Mask R-CNN (3) model to automatically segment knee joints and differentiate the sides for report generation. These segmented knee joints were further used to train regression and classification models based on DenseNet-121(4) to automatically predict the KL-grade from an image. Both regression and classification models trained on the OAI were fine-tuned using the Target dataset, and comparison of performance of both models on both the dataset was done.

The regression model outperformed the classification model. It correctly classified 73.26% of the cases in the Target set where 87.35% of the misclassifications were amongst neighbouring KL grades. Mean absolute error of 0.2817, precision of 0.73, and recall of 0.73 was obtained.

Regression considers the inherent ordering between the KL grades and significantly decreases misclassification in non-adjacent classes as compared to classification. Automated and accurate assignment of KL grades to knee radiographs, can help mitigate the effects of radiologist subjectivity in assessment; reducing radiologist workload, and improving reporting times. Also, one single model is unlikely to work across different cohorts but can be finetuned to the concerned cohort.