Deep convolutional neural networks for age and gender estimation using orthopantomogram images: a systematic review

The primary goal of this systematic review is to critically analyze and evaluate how effectively deep convolutional neural networks (CNNs) can estimate age and gender from orthopantomogram (OPG) images. OPG images, commonly known as panoramic dental X-rays, provide a comprehensive view of the upper and lower jaws, teeth, and surrounding structures, making them valuable for a variety of dental and forensic applications. The objective is to provide a comprehensive evaluation of the efficiency of CNNs in this specific application, offering valuable information for researchers and practitioners interested in leveraging advanced AI techniques in dental imaging.

The protocol is registered in PROSPERO before data collection (CRD42024578905). A comprehensive literature search was conducted across multiple databases, including Research Scholar, PubMed, Research Gate, Springer and Science Direct, focusing on studies published between [2014 to 2024]. The inclusion criteria were studies employing CNNs for age and gender estimation using OPG images. Data were extracted on study design, CNN architecture, sample size, and performance metrics, followed by a comparative analysis of the outcomes. The risk of bias assessment used the latest version of the Quality Assessment tool for Diagnostic Accuracy Studies (QUADAS-2).

The review initially identified 194 studies focusing on various CNN architectures for age and gender estimation. After removing 25 duplicate records, 169 unique articles remained. Further screening narrowed these down to 7 studies that met the criteria for inclusion in the systematic review. The findings suggest that CNN-based methods typically surpass traditional techniques in terms of accuracy and reliability. However, the results varied widely across studies due to factors like dataset size, image quality, and the complexity of the CNN models used. Most studies were conducted in single centers and only 20% included an external test set for evaluating model performance. According to the QUADAS-2 AI tool, only 40% of the studies in this review were deemed to have a low risk of bias in the reference standard domain. COMPARATIVE ANALYSIS OF CNN ARCHITECTURES AND REASONS FOR PERFORMANCE DIFFERENCES: Several studies of OPG images experimented with multiple CNN architectures including VGGNet, ResNet, DenseNet, Inception, MobileNet, and even custom models; each one helping in a different way to extract the features from the OPG images. All these networks have different depth, connectivity, and efficiency, which thus directly influenced their ability to reveal the subtle dental and maxillofacial features that are important for age and gender estimation. In particular, VGGNet provided excellent low-level feature extraction but was also the most prone to overfitting, whereas architectures like ResNet and DenseNet were always in the lead owing to the residual and dense connections that facilitate gradient flow, boost stability and enhance feature reuse. The Inception models were winners mainly because of their ability to extract features at different scales, and MobileNet was the next best option in terms of performance but only at the cost of having less depth in representation. As a result, models built on ResNet and DenseNet were the most generalizable and accurate, hence the best in spotting the subtle developmental and morphological characteristics in OPG images.

Deep CNNs demonstrate significant potential for age and gender estimation from OPG images, but further advancements in data collection and model refinement are necessary to enhance their accuracy and generalizability. This review provides a foundation for future research aimed at improving CNN-based methods in dental imaging. It demonstrates wide variety of methods used. Including segmentations, classification, Pre-processing and labeling techniques.