Dental image analysis and surgical decision support using modified generative adversarial networks

Several clinical issues have remained in dental care and risk estimation because of the insufficient integration of multimodal features, the incapacity to integrate clinical domain literatures and evidence-based principles in their analytical platforms. The methods of deep learning-based diagnosis offer the in-depth examination of complex cases synthesized based on heterogeneous sources, hence, they require appropriate structural adjustments that can be implemented with the innovative components into a single system.

A modified Generative Adversarial Network (mGAN) is presented, which incorporates three key innovations: (1) We introduce a novel W-Net architecture which is formed by combining U-Net and V-Net to develop dual-path encoding and skip-gated decoding mechanisms. It enables multi-scale feature analysis capabilities specifically optimized for complex dental anatomical structures. (2) Large Language Model (LLM) supports to incorporate hyper parameter turning of acquired refined features based on conditional and guided generator based on domain knowledge distilled on mandibular anatomy and prevalent dental pathology directly into the deep learning pipeline. Thus, it shows complex medical scenarios beyond purely data-driven pattern recognition. (3) Probability Feature Fusion (PFF) is a novel mechanism for adaptive multi-modal feature integration, enabling dynamic weighting of different information sources based on local image content and clinical context from LLM. PPF combines multi-modal features within a unified analytical framework for maintaining interpretability and clinical relevance for generating probabilistic ground-truth maps representing contact points between mandibular third molar (M3M) roots and inferior alveolar canal (IAC).

The mGAN framework proposed had 95.7% Root Surface Area Accuracy (RSAA) in a clinical assessment of dental and a Dice Similarity Coefficient (DSC) of 0.925 in the segmentation of anatomical structures. The system had a sensitivity of 91.8% and specificity of 96.7% of surgical necessity classification relative to expert consensus. Ablation experiments inferred that the knowledge integration with LLM improved risk correlation by 6.4% over baseline GAN models, and Probability Feature Fusion added an additional 1.9% improvement in sensitivity. All the improvements reported with respect to the closest baseline (TSTNet) are similar everywhere and non-overlapping 95% confidence intervals on central measures. The time of processing was 2-2.2 min per case, which is appropriate to be included in clinical workflows. We tested the method using 1500 CBCT scans that were annotated by experts.

The proposed mGAN architecture can combine domain knowhow with deep learning in dental surgical planning. This approach can offer a potential tool to support clinical decision-making in oral surgery by offering uncertainty-sensitive predictions and preserving computational efficiency, which has the potential to minimize the iatrogenic complications and enhance the treatment outcomes.