使用生成式對抗網路之牙齒色階與形狀預測系統

Abstract

臨床牙科比色一直是牙醫師與技師所面臨的挑戰。傳統比色方法依賴人工肉眼判斷，正確率不高，常導致牙貼重作，造成時間與成本損失。市面上的比色輔助設備價格高昂且需專用儀器，此外，牙齒表面的反光也會影響比色準確性。 本研究提出一套深度學習牙齒比色系統，整合牙貼預測與反光去除兩大功能。透過暗箱與偏光攝影建立資料集，並以生成對抗網路訓練兩組模型：一組預測牙貼套用效果，另一組去除鏡面反光。牙貼預測模型為本研究首次提出，能根據支台齒外觀生成擬真牙貼影像，具高度應用潛力，並透過資料擴增提升其在不同角度、色階與厚度下的穩定性。實驗顯示，本系統的反光處理效能可與現有通用型深度學習模型相當，並優於傳統影像修復技術。系統操作簡便，僅需相機與預訓練模型即可應用於臨床，協助快速直觀預測牙貼效果，提升溝通效率並降低成本。

Shade matching in dentistry is a long-standing challenge for dentists and technicians. Traditional visual methods are often inaccurate, leading to veneer remakes, wasted time, and higher costs. Existing tools are expensive and require specialized hardware. Reflections on tooth surfaces also interfere with accurate shade assessment. This study presents a deep learning–based system that combines veneer prediction and reflection removal. Using a darkroom and polarized photography, we built two datasets and trained two Generative Adversarial Network models. One predicts realistic veneer appearances based on abutment teeth, while the other removes specular reflections. The veneer model is newly proposed and enhanced with data augmentation for better performance under various conditions. Results show that our reflection removal model performs comparably to general deep learning models and outperforms traditional image-based techniques. The system is easy to use, requiring only a camera and pretrained models, and supports fast, intuitive veneer simulations to improve communication and reduce clinical costs.