Thank you for that introduction and AOSSM for inviting me to speak today and my co-authors. I don't have any disclosures. So imaging is ubiquitous in medicine and orthopedics is no exception to that. As we've discussed today, deep learning has its applications for medical imaging analysis in part for surgical planning and diagnosis. And as we've highlighted today, some of the important requirements for machine learning is large data and, as noted earlier, external validation, which is what I found exciting about this publication last year of the RAD ImageNet database. This is a database of 1.4 million images made up of CTs, MRIs, and ultrasounds from 11 different anatomic locations gathered from over 130,000 patients. And these images were labeled and annotated by board-certified fellowship-trained radiologists. And the investigators who compiled this database looked at and were able to develop their own deep learning models using these images and were able to accurately identify ACL and meniscal tears. However, looking at rotator cuff pathology in this database, although they had thousands of shoulder MRIs, is currently unknown. So that's what we attempted to look at. We wanted to differentiate rotator cuff pathology from normal anatomy on shoulder MRIs within this dataset using deep learning. So what we did is we took all of the labeled supraspinatus tears in that dataset, both partial and full thickness, as well as tendinoses. We took all those images, and those were already relabeled by the compilers of the dataset, and the images were at the site of most importance for identifying that pathology. And we took all of the normal shoulder MRIs as well. In the normal group, we excluded all images medial to the glenoid and in the sagittal plane and all axial views to better approximate the data in the pathologic group set. This resulted in a total of 6,173 images, 85% of which represented supraspinatus pathology and 15% of which represented normal shoulder anatomy. This is a flow diagram of our study design. And I'd like to just draw attention to kind of the split between training, validation and testing data that was split in an 80%, 10% and 10% fashion respectively. We set aside the testing data so that way it was never seen during the training process. During the training process we used Convinex deep learning model as our base architecture for transfer learning. And after training our model, we had a final trained model that we used on this novel set of MRIs to make our final predictions. What we found is that this model had quite good results with an accuracy of 98.9%, a sensitivity of 98.7% and a specificity of 100%. This is a heat map showing the relative importance for our model. Higher areas of importance are shown in yellow and those of less importance are shown in purple. And what we see is that the area that is most important for our model making its prediction is the area of the insertion of the supraspinatus tendon on the humeral head. So looking at a few other papers with similar methodology, we find that deep learning performs similarly well in classifying supraspinatus tear and rotator cuff pathology. Shim for example had an accuracy of 92.5% looking at roughly 2,000 patients in their MRIs and Yao had an AUC of 0.94 when looking at their data set of 200. So as we've talked about today, there are many potential applications of AI based technology in clinical practice. However, one of the limitations to adopting these technologies is the need for external validation and that's because, you know, collecting these large data sets is labor intensive and being HIPAA compliant presents some barriers as well. So this RAD ImageNet database with a lot of images could be used for external validation in the future. Some limitations of our current study is that our model performance could not be stratified by the severity of the supraspinatus injury and our models were trained on static two-dimensional images when MRI is three-dimensional capturing of, you know, this three-dimensional anatomy and we may be losing some information in that. And as our study was primarily focused on the internal validity of the RAD ImageNet database, external validation was not performed. So in conclusion, supraspinatus pathology can be reliably differentiated from normal anatomy using deep learning. Deep learning models demonstrate excellent accuracy and predictive value and RAD ImageNet database may serve as a valuable resource for external validation in the future. Thank you.

deep learning

medical imaging analysis

orthopedics

RAD ImageNet database

rotator cuff pathology