Purpose: The early detection and diagnosis of rare genetic diseases are crucial for maintaining patient health and well-being. However, the diagnosis of said diseases can be challenging owing to their rarity, limited clinical experience of physicians, and specialized facilities required for diagnosis. In recent years, deep learning algorithms have been investigated as a potential strategy for the efficient and accurate diagnosis of these diseases. Herein, we used the deep learning algorithm of multitask cascaded convolutional neural networks (MTCNN) to develop a face recognition model for the diagnosis of velocardiofacial syndrome (VCFS).
Methods: We trained the model on a publicly available labeled face dataset and evaluated its performance. Subsequently, we analyzed the binary classification performance of diagnosing VCFS using the most efficient face recognition model. The facial images of 98 patients with VCFS and 91 non-VCFS controls who visited Seoul National University were used to create training and test sets. Moreover, we analyzed whether the classification results matched the known facial phenotype of patients with VCFS.
Results: The facial recognition model showed high accuracy, ranging from 94.03% to 99.78%, depending on the training dataset. In contrast, the accuracy of the binary classification model for the diagnosis of VCFS varied from 80.82% to 88.02% when evaluating with photographs taken at various angles of the patient depending on the structure. When only evaluating frontal photographs, the accuracy was 95.00%. Moreover, the importance level analyzed through the gradient-weighted class activation mapping heatmap showed the characteristic parts of perinasal and periorbital area to be consistent with the conventional facial phenotypes of VCFS.
Conclusion: We attempted to diagnose the patients' genetic syndrome using MTCNN-based deep learning algorithms only with the photos of the faces of patients with VCFS. We obtained high accuracy, and deep learning-based diagnosis has been conformed to be greatly helpful for medical staff in the early detection and diagnosis of children with rare genetic diseases, enabling them to provide treatment in a timely manner.