Abstract
Frank's sign (FS) is a diagnostic marker associated with aging and various health conditions. Despite its clinical significance, there lacks a standardized method for its identification. This study aimed to develop a deep learning model for automated FS detection in 3D facial images derived from MRI scans. Four deep learning architectures were evaluated for FS segmentation on a dataset of 400 brain MRI scans. The optimal model was subsequently validated on two external datasets, comprising 300 brain MRI scans each with varying FS presence. Dice similarity coefficient (DSC) and receiver operating characteristic (ROC) analysis were employed to assess model performance. The U-net architecture demonstrated superior performance in terms of accuracy and efficiency. On the validation datasets, the model achieved a DSC of 0.734, an intra-class correlation coefficient of 0.865, and an area under the ROC curve greater than 0.9 for FS detection. Additionally, the model identified optimal voxel thresholds for accurate FS classification, resulting in high sensitivity, specificity, and accuracy metrics. This study successfully developed a deep learning model for automated FS segmentation in MRI scans. This tool has the potential to enhance FS identification in clinical practice and contribute to further research on FS and its associated health implications.