Measuring the Accuracy of Artificial Intelligence Algorithms in the Identification of Anatomical Landmarks Through the Rastereography SPINE3D

Abstract

Rasterstereography represents a valid alternative as a screening tool for analyzing spinal anomalies, thanks to its numerous advantages over the invasiveness of radiological techniques. The validity and reliability of rasterstereography are still questioned, especially when considering the accuracy of the artificial intelligence algorithm for the identification of the anatomical landmarks. To validate the commercial system SPINE3D, tests were conducted on ten healthy subjects, instructing them to maintain a relaxed standing position. The tests were repeated five times. To proceed with the validation of SPINE3D, the points identified by SPINE3D were compared with those acquired through an optoelectronic system, used as the gold standard. The measurements allowed to assess the accuracy of artificial intelligence algorithms for detecting anatomical landmarks, also considering the uncertainty due to the repeatability of the measurement process. The maximum absolute error was found always lower than 3.7 cm, and in average ranged from 1.7 to 2.9 cm. The anatomical landmarks related to the upper side of the back seems to be the ones affected by a greater error in the estimation; specifically, the prominence corresponding to the seventh cervical vertebra was associated with an absolute error ranged from 2.0 and 3.4 cm, Further analyses must be conducted to understand how the accuracy in the estimation of landmarks influences the synthetic indices computed for the assessment of spinal abnormalities.

Ähnliche Arbeiten