Automated Sella-Turcica Annotation and Mesh Alignment of 3D Stereophotographs for Craniosynostosis Patients Using a PCA-FFNN Based Approach.

BACKGROUND

Craniosynostosis, characterized by the premature fusion of cranial sutures, can lead to significant neurological and developmental complications, necessitating early diagnosis and precise treatment. Traditional cranial morphologic assessment has relied on CT scans, which expose infants to ionizing radiation. Recently, 3D stereophotogrammetry has emerged as a noninvasive alternative, but accurately aligning 3D photographs within standardized reference frames, such as the Sella-turcica-Nasion (S-N) frame, remains a challenge.

METHODS

This study proposes a novel method for predicting the Sella turcica (ST) coordinate from 3D cranial surface models using Principal Component Analysis (PCA) combined with a Feedforward Neural Network (FFNN). The accuracy of this method is compared with the conventional Computed Cranial Focal Point (CCFP) method, which has limitations, especially in cases of asymmetric cranial deformations like plagiocephaly. A data set of 153 CT scans, including 68 craniosynostosis subjects, was used to train and test the PCA-FFNN model.

RESULTS

The results demonstrate that the PCA-FFNN approach outperforms CCFP, achieving significantly lower deviations in ST coordinate predictions (3.61 vs. 8.38 mm, P <0.001), particularly along the y -axes and z -axes. In addition, mesh realignment within the S-N reference frame showed improved accuracy with the PCA-FFNN method, evidenced by lower mean deviations and reduced dispersion in distance maps.

CONCLUSIONS

These findings highlight the potential of the PCA-FFNN approach to provide a more reliable, noninvasive solution for cranial assessment, improving craniosynostosis follow-up and enhancing clinical outcomes.