CNN, GAN, and RNN Architectures for Anatomical Segmentation, Surgical Planning, and Outcome Prediction in Combined Orthognathic and Maxillofacial Surgery: A PRISMA-ScR Scoping Review

OVERVIEW- Deep learning architectures, with emphasis on the foundational pipelines- CNN, GAN and RNN are explored in this project across combined orthognathic and maxillofacial surgical workflow, under the protocol and in accordance with PRISMA ScR framework. BACKGROUND- The procedures involved in combined orthognathic and maxillofacial surgery are technically challenging. Craniofacial diagnosis, 3D virtual surgical planning, osteotomy design and simulation, intraoperative execution, and long-term post-operative evaluation of skeletal stability, soft tissue adaptation, airway dimensions, and functional results are all integrated into these processes. High levels of spatial thinking, anatomical precision, and prediction accuracy are required for each of these stages, which have traditionally relied on clinical knowledge and are highly variable amongst operators. The field of surgical decision assistance and medical image analysis has completely changed since deep learning emerged as a revolutionary subset of AI. Particularly in the field of craniofacial surgery, three neural network families have proven to have the most reliable and useful applications: CNN-The automatic segmentation of anatomical structures from computed tomography (CT), cone beam CT (CBCT), and MRI datasets has made substantial use of Convolutional Neural Networks (CNNs), including designs like U-Net, ResNet, VGG, and DenseNet. These include segmentation of cortical and cancellous bone, mandibular and maxillary structures, soft tissue envelopes, the inferior alveolar nerve, and pathological lesions, including tumours and cysts. Additionally, CNN-based landmark detection algorithms have proven effective for automated cephalometric analysis, enhancing consistency and reducing planning time. GAN-In order to solve the ongoing issue of small dataset sizes in uncommon craniofacial conditions, Generative Adversarial Networks (GANs), such as CycleGAN, pix2pix, and their derivatives, have introduced the ability to generate photorealistic pre- and post-surgical simulations, perform image-to-image translation between imaging modalities, and synthesise training data. In an area where traditional mathematical models have historically performed poorly, GAN-based methods have demonstrated significant promise in producing patient-specific soft tissue forecasts following Le Fort osteotomies and bilateral sagittal split osteotomies (BSSO). RNN-The ability to model temporal and sequential data is introduced by RNNs, especially LSTM and Gated Recurrent Unit (GRU) architectures. This capability is particularly pertinent to surgical outcome prediction. Applications include airway volume changes after maxillomandibular advancement, longitudinal soft tissue remodelling, post-operative skeletal relapse trajectories, and functional outcome sequences like speech and swallowing recovery. No published scoping review has thoroughly mapped the evidence on all three design families across the entire combined orthognathic-maxillofacial surgical process, despite a fast growing body of primary literature. Previous studies have either focused on a single workflow stage, like planning or segmentation, or they have been procedure-specific or architecture-specific. As a result, there is no unified evidence map to help physicians, academics, and health technology organisations make decisions about clinical translation, financing goals, or research agendas.