Deep Learning-Based Automated Detection of Pediatric Dental Lesions: Combining Swin-Transformer and Mask Region-Based Convolutional Neural Networks

Access to pediatric dental care is often restricted by a lack of specialized dentists and the high cost of diagnostics, leading to delayed treatment. This study introduces a deep learning methodology for identifying pediatric or deciduous dental lesions in clinical images, offering two significant advancements. First, we developed a pediatric dental lesion dataset comprising 390 high-resolution images annotated with 2,114 detectable dental pathologies, ranging from small to moderate to severe lesions/tooth decay, which provides comprehensive resources for studying pediatric dental lesions. Second, we propose a novel deep learning framework that combines the Swin-Transformer with Mask region-based convolutional neural networks (R-CNNs) for precise lesion detection. This model is adept at identifying signs of various pediatric dental lesions, including patients with multiple disease sites. The Swin-Transformer component extracts features from image patches, capturing both local details and broader contextual information. Mask R-CNN then uses these extracted features to accurately identify and differentiate between cooccurring lesions within the same image. This approach addresses the challenge of diagnosing multiple lesion types that are present simultaneously, a common scenario in pediatric dentistry. To our knowledge, these contributions represent a novel approach for the detection of pediatric dental lesions. The testing of our model on the proposed dataset shows substantial improvements in both the accuracy and efficiency of image-based diagnoses, validating the effectiveness of our proposed method.