Summary

An unsupervised machine learning model capable of detecting meniscal injuries and anterior cruciate ligament injuries

Abstract

Objective

To establish a predictive method for the detection of meniscal and/or anterior cruciate ligament injuries through an artificial intelligence model

Method

1250 non-contrast MRI images of the knee in T1 and T2 sequences were analyzed in their axial, coronal and sagittal sections. The images were extracted from the open source database of Stanford University, USA. They were manually reviewed and classified according to the injuries found in each case, tabulating between meniscal injuries, anterior cruciate ligament injuries and other injuries. Meniscal injuries were sub-classified according to topographic area and type of injury. The images and data tabulation were processed by a Self-Supervised Learning for Knee Injury Diagnosis predictive artificial intelligence model (SKID Model). Receiver Operating Characteristic (ROC) curves for the prediction of meniscal injury and anterior cruciate ligament injury were obtained without subcategorizing meniscus or other injuries.

Results

Under the SKID model, a ROC curve of 0.90 was obtained for the detection of abnormal images, 0.89 for the detection of anterior cruciate ligament injuries, and 0.80 for the detection of meniscal injuries.

Discussion

According to the current literature, the use of the SKID artificial intelligence model for the detection of abnormal images in the medical field has not been reported. Among the artificial intelligence models published to date, no ROC curves equal to or greater than 80% have been reported for meniscal injuries and/or over 89% for anterior cruciate ligament injuries, so new predictive subclassifications will continue to be investigated with this model for the detection of specific injuries in knee images.

Conclusion

Of the predictive artificial intelligence models, the SKID model offers attractive results to continue deepening its use and study for the detection of knee injuries through the use of images.