MRI can be used to assess several structural changes related to knee OA (KOA). Software image analysis methods provide objective fully quantitative measurements of these changes. However, such methods generally require a human reader and can be time consuming if the number of images in a study is large. Deep leaning (DL), a form of statistical machine learning, offers the potential for increased or full automation, therefore reducing reader time substantially.
Our U-Net based DL approach was trained on 600 subjects from the Osteoarthritis Initiative (OAI) dataset. We used the baseline and 24 month scans to train a Bone Marrow Lesion (BML) detector, given the MRI images and semgentation masks annotated by trained radiologists, obtained using a previously validated semi-automated (SA) method.
|Figure 1. Example of a BML in the patella on MRI (left) and segmented (right). Yellow pixels are where DL and SA overlapped. Red and green pixels are areas of disagreement.|
The average Dice similarity coefficient between predicted BML segmentation masks and human annotated segmentation masks was 0.85. The Pearson’s R2 coefficient was very high (0.99).
The results suggest that the SA and DL methods are nearly equivalent for segmenting BMLs for KOA. The DL approach can produce a highly accurate method to assess BML volume that offers substantial time savings over the SA method.
F. Preiswerk, M. Sury, J. Wortman, J. Collins, and J. Duryea, “Application of Deep Learning for the Quantitative Assessment of Bone Marrow Lesions (BMLs),” in Proceedings of the International Workshop on Osteoarthritis Imaging, Menton, France, 2018.