Dr. Charles Fang from Santa Clara Valley Medical Center in San Jose, CA, and researchers from Stanford University developed a two-part machine-learning process that directs a regional convolutional neural network to zero in on the arcuate region of the knee to detect Segond fractures, which are frequently associated with ACL tears.
The researchers reviewed 235 frontal knee radiographs and recorded cases of severe arthrosis or deformities around the arcuate region. Next, they trained the CNN to identify the arcuate region using 172 normal knee x-rays. They set aside the remaining 63 x-rays to test the CNN's accuracy in discerning the abnormalities and their locations.
To further train the regional CNN to classify injuries, the group implemented a second phase with only 24 frontal knee x-rays with Segond fractures and 129 normal knee x-rays. The researchers also created a separate test set of x-rays that included five Segond injuries and 15 normal cases.
Results from the first part of the study showed that the regional CNN was accurate in detecting and localizing anomalies in the arcuate region in all 63 test cases (100%), meaning that the "expert level" of diagnosis was achieved for Segond fractures and validation of an ACL tear. In the second phase, the CNN was 100% accurate in classifying the Segond fractures and normal x-rays.
Moreover, the results were achieved with an "extremely sparse set of training data (24 samples)," Fang and colleagues wrote. CNNs "may address the challenge of big data in radiology machine learning by reducing the size of training sets required to teach AI systems to diagnose subtle and rare pathology with high sensitivity and specificity."
