Application of a neural network to restore the lost surface of skull bones

Objective: to evaluate the sensitivity, specificity and accuracy of a digital algorithm based on convolutional neural networks to restore of bones of cranium defects. Material and methods. Neural network training was carried out as a result of 6,000 epochs on 78,000 variants of skull models with artificially generated skull injuries. The evaluation was performed on 222 DICOM series of patients computerized tomography with bones of cranium defects. Results. The indicators of sensitivity, specificity and accuracy were 95.3%, 85.5% and 79.4% respectively. A number of experiments were carried out with step-by-step sorting of three-dimensional models in order to find the reasons for the unsatisfactory skull reconstructing results. Incorrect detection of the skull defect most often occurred in the area of the facial skeleton. After excluding the series with artifacts, the average increase in metrics was 2.6%. Conclusion. Correct determination of the bone defect at the scull model (specificity) by the algorithm had the greatest impact on the surface accuracy. The maximum accuracy of the algorithm, which allows using the obtained surfaces without additional processing in a three-dimensional modeling environment, was achieved on series without the presence of artifacts during computed tomography (83.5%), as well as with defects that do not extend to the skull base (79.5%).