pix2pixHDv2: Efficient Panoramic Dental Image Synthesis with Minimal Artifacts and Computational Overhead

Karacan M. H., Yücebaş S. C.

JOURNAL OF IMAGING INFORMATICS IN MEDICINE, ss.0-17, 2026 (SCI-Expanded)

Özet

Although recent image-generation studies have achieved promising results, they still present challenges in terms of computational cost and training/inference time that need to be addressed. Conditional generative adversarial networks (mask-to-image) such as pix2pixHD can synthesize high-resolution panoramic dental radiographs, yet they often introduce underreported artifacts (e.g., circular blotches and checkerboard patterns). This study characterizes these artifacts and proposes pix2pixHDv2, a lightweight architecture that replaces ResNet residual blocks with ConvNeXt and ConvNeXtV2 modules and substitutes transposed-convolution upsampling with bilinear upsampling or PixelShuffle. Using five public datasets (OdontoAI, TUFTS, ADLD, TSXK, AKUDENTAL), models were trained under matched hyperparameters and evaluated for distributional similarity and perceptual quality. Gradient-weighted Class Activation Mapping (Grad-CAM) shows that the proposed residual design shifts early attention toward anatomically relevant tooth regions and suppresses spot-like artifacts, while frequency-spectrum analysis indicates reduced checkerboard periodicity. These findings demonstrate that targeted architectural refinements can substantially improve both image fidelity and resilience against artifacts in dental imaging applications. The ConvNeXtV2 + ConvTranspose2d configuration achieved improvements of 16.3%, 36.5%, 2.8%, and 5.1% in FID, KID, LPIPS, and DISTS, respectively. In addition, synthetic images generated using different backbone and upsampling configurations were presented to 16 specialists in oral and maxillofacial radiology, and an AUROC of 0.65 indicated that distinguishing synthetic images from real radiographs was challenging for expert observers. The addition of synthetic images was associated with a 5.2% relative increase in segmentation performance. Code is available at: https://github.com/merterhk/pix2pixHDv2