NAADA: A noise-aware attention denoising autoencoder for dental panoramic radiographs

Low-dose radiation imaging of dental panoramic radiographs often produces a complex noise pattern that significantly degrades image quality, thereby limiting diagnostic utility. This creates a need for advanced post-processing methods capable of extracting core anatomical structures from noisy regions. While convolutional denoising autoencoders (DAEs) are powerful tools for image restoration and have shown notable success in denoising tasks, they inherit a key limitation of convolutional neural networks (CNNs): they tend to recover low-frequency features, such as smooth regions, more effectively than high-frequency details. This often results in the loss of fine structures, which is particularly problematic in dental radiographs where preserving subtle anatomical features is critical. While self-attention mechanisms can help mitigate this issue by emphasizing important features, conventional attention methods often prioritize cleaner regions and may overlook features obscured by noise. To address this limitation, we propose a noise-aware self-attention mechanism that enables the model to focus on and recover key features even within noisy regions. Building on this concept, we introduce the noise-aware attention-enhanced denoising autoencoder (NAADA) network for enhancing noisy panoramic dental radiographs. Compared with recent state-of-the-art but heavier models, such as Uformer and MResDNN, our method improves the reconstruction of fine details, ensuring superior image quality and diagnostic reliability.