Intern / Post-doctoral Research Fellow Parkland Memorial Health / UT Southwestern Medical Center Dallas, Texas
Abstract: Problem Statement: Cleft lip and palate (CLP) are of the most common defects treated by surgeons with a reported prevalence of 1 in every 1600 babies in the United States. Accurate identification of the alveolar cleft site with volumetric analysis provides a foundation for more accurate surgical planning and improved outcomes. Traditionally, 2D imaging modalities were the standard for defect analysis; however, the advent of 3D imaging modalities revolutionized surgical planning. Recent advancements in artificial intelligence and deep learning provide a promising avenue for highly accurate surgical planning. Accurate calculation of alveolar cleft volumes remains a challenging concept due to subjectivity and variation of landmark identification by providers. Convolutional Neural Networks (CNNs) have been previously utilized for identification of both anatomical structures and pathology in radiology with high levels of accuracy; however, their utilization remains limited in craniofacial surgical planning. This study demonstrates the utilization of a modified Residual Neural Network (ResNet) model for calculation of alveolar cleft volumes on head and neck CT imaging.
Material and
Methods: A total of 52 deidentified CT datasets from patients with unilateral alveolar clefts were obtained and converted to DICOM files. Inclusion criteria included CT imaging data from unilateral CLP patients with an alveolar cleft planned for subsequent SABG and in mixed dentition at the time of imaging. Previously described alveolar cleft landmarks by Linderup et al. were utilized to identify the margins of the defect in buccal/palatal, mesiodistal and superior-inferior planes of the axial slices. CT datasets were manually annotated by experienced craniofacial surgeons. Annotations were subsequently utilized via transfer learning for training of a modified ResNet-152 deep learning model on PyTorch. 90.4% (N=47) datasets were utilized for training of the model followed by a subsequent 3.8% (N=2) for validation and the remaining 5.8% (N=3) for evaluation of model performance.
Data Analysis and
Results: For evaluation of model performance, each scan data was independently analyzed by surgeons (gold standard) and the ResNet model. Segmented 2D data from each group's analysis was then used to create a 3D mask and the volume of alveolar cleft was calculated. The calculated alveolar cleft volume was recorded from the ResNet model and surgeon manual annotation sets and volumetric values were analyzed via paired t-test on SPSS to determine statistical significance (P-value < 0.05) The mean alveolar cleft volume was calculated as 0.97 cm3 and 0.72 cm3 for the ResNet model and surgeon datasets accordingly with no statistically significant difference observed between the groups (p=0.254).
Conclusion and Outcomes Data: The modified ResNet-152 deep learning model utilized in this study was trained and accurately identified landmarks for volumetric analysis of alveolar clefts similar to those observed from manual annotations by surgeons. Limitations of the study at the current state included the imaging data available and restrictions in computational capacity for adequate model training and verification. While basic at its current state, the data obtained from the model provides an introduction to utilization of machine learning in oral and maxillofacial surgery for surgical planning and illustrates the limitless opportunities in the future. Such models can be further trained and modified to perform similar tasks in various aspects of the specialty; Thus, providing an accurate, automated and highly individualized modality for surgical planning and enabling surgeons to increase intraoperative efficiency and limit discrepancies between imaging and clinical findings.
Liu, Bing, et al. "A novel accurate volumetric analysis protocol for evaluating secondary alveolar cleft reconstruction." Journal of Cranio-Maxillofacial Surgery 48.7 (2020): 632-637.
Linderup, Bo Werner, et al. "A novel semiautomatic technique for volumetric assessment of the alveolar bone defect using cone beam computed tomography." The Cleft Palate-Craniofacial Journal 52.3 (2015): 47-55.
