Deep-learning models can be used to grade the extent and severity of osteoarthritis in the knee1. The disease occurs when cartilage cushioning the ends of bones in the knee joint gradually wears down.
The models can decode hidden features of the disease in histology images and will aid clinicians in assessing cartilage damage, says a team at the Indian Institute of Technology Delhi.
Pathologists manually grade osteoarthritis-related histology images. This process is cumbersome and prone to error. Scientists, led by Sourabh Ghosh, trained four deep-learning models to grade histology images depending on three parameters – staining intensity, cell distribution and arrangement, and cell morphology.
The team, which included Raju Vaishya at the Indraprastha Apollo Hospital in New Delhi, say these parameters led to a scoring system ranging from 0 to 9. This can be further divided into four grades – 0 to 4 – where 0 indicates a healthy cartilage and 4 reveals a severely damaged cartilage.
Of the models, DenseNet121 is the most efficient, yielding a grading accuracy of 84%. Another model, ResNet50, showed an accuracy of 81%.
These models could outperform and replace human-supervised analysis, eliminating its ambiguity in grading cartilage damage, says Ghosh.