Can a cephalogram be used to estimate palatal suture maturity to inform expander selection

The maturity of the mid-palatal suture (MPS) could largely determine the necessity of more invasive interventions for correcting transverse maxillary deficiency. CBCT, the most commonly used method for assessing MPS maturity, is not routinely available and exposes patients to higher levels of ionizing radiation than a cephalogram. Given the substantial number of morphological features contained in a cephalogram, this study aimed to estimate CBCT-defined MPS maturity using an entire cephalogram alone via deep learning (DL) algorithms.

A total of 543 paired CBCT and cephalogram files available from two distinct hospital settings formed the internal and external datasets. Data augmentation and five-fold cross-validation strategies were applied. Six ResNet-based DL algorithms were initiated via transfer learning. The AdamW optimizer and an early-stopping strategy were employed during training. The models were subsequently tested on the held-out external dataset. The Grad-CAM technique was applied to offer insights into the models' decision-making process.

All algorithms exhibited promising classification performance, with area under the curve (AUC) values ranging from 0.91 to 0.97. The models with the TKSGA module demonstrated better discriminative ability, calibration, and clinical usefulness than their baseline counterparts. Models integrated with the TKSGA module exhibited an interpretable decision pattern. The heatmaps revealed that morphological characteristics of patients' dentition, cervical vertebrae, and nasopharyngeal airway were primary areas of interest for the DL models.

When CBCT imaging is unavailable or not justified, a lateral cephalogram coupled with advanced DL algorithms, could be considered as a supplementary tool to inform palatal expander selection.