A new study to be published in Spine has found that deep learning applications are able to identify previous spinal implants. In some cases, Hee-Seok Yang et al report over 90% precision using deep learning.
The paper, entitled “Deep learning application in spinal implant identification,” is a retrospective observational study of clinical cases in which patients had lumbar spine one-segment instrument surgery. Researchers used nearly three thousand radiographs, 1,446 from the lumbar spine anteroposterior (AP) and 1,448 from the lumbar spine lateral. Five different implants were labelled in this study, all of which were different pedicle screws.
These 2,895 radiograph images of implants were then fed into three different deep learning algorithms: the traditional deep neural network model built by coding the transfer learning algorithm, Google AuthoML, and Apple Create ML. The algorithms’ effectiveness was measured on recall and precision.
According to the researchers, “despite many attempts to apply deep learning to medical images, the application has rarely been successful.” Authors report that the goal of this study was to demonstrate the clinical usefulness of deep learning by using this tool to identify previous spinal implants in radiographs.
Researchers report that over all the models performed “well”. The conventional transfer learning when used on the AP images showed 97% precision and 96.7% recall. On the lateral radiography it showed 98.7% precision and 98.2% recall.
Researchers claim that Google AutoML showed 91.4% precision and 87.4% recall for AP radiography. In the case of the lateral radiography images the Google AutoML showed 97.9% precision and 98.4% recall.
Finally, authors report that in the case of Apple Create ML, the programme showed 76% precision and 73% recall for AP radiography. In the case of lateral radiography, this algorithm showed 89% precision and 87% recall.
The authors noted that, “in all deep learning algorithms, precision and recall were higher in lateral than in AP radiography.”
The researchers concluded that, “the deep learning application is effective for spinal implant identification,” and further commented that clinicians can use machine learning based deep learning applications to improve clinical practice and care.