基於合成條紋渲染器的無監督領域自適應之於口腔內圖像結構光中心線提取

Abstract

口腔內結構光（SL）成像因口腔環境中表面複雜且高反射性，常面臨中心線品質下降的挑戰。雖然完全監督式學習方法可以改善中心線檢測，但它們需要勞動密集型的標註工作，使其在大規模數據集上變得不切實際。為了解決這些挑戰，我們提出了一個合成條紋渲染器（Synthetic Stripe Renderer，SSR）：以常規的口腔內圖像作為輸入，SSR 利用 Segment Anything 生成的區域掩模來合成與隨機條紋圖案配對的口腔內 SL 圖像。我們將 SSR 與無監督領域適配（UDA）訓練集成到 U-Net 模型中，實現無需人工標註或配對領域數據的從合成到真實的中心線提取。此外，我們提出了precision@d、recall@d 和 f1-score@d等指標，用於通過中心線坐標集的重疊率來評估多條紋段的相似性。我們的方法能夠提取高品質的中心線區域，其性能可與全監督式訓練相媲美。

Intraoral structured light (SL) imaging often faces challenges with degraded centerline quality due to the complex and reflective surfaces in the oral environment. While fully supervised learning approaches can improve centerline detection, they require labor-intensive annotation, making them impractical for large-scale datasets. To address these challenges, we propose a Synthetic Stripe Renderer (SSR): Given regular intraoral images as input, SSR employs region masks generated from Segment Anything to synthesize intraoral SL images paired with random stripe patterns. We integrate the SSR and unsupervised domain adaptation (UDA) training into a U-Net model, enabling synthetic-to-real centerline extraction without requiring manual labels or paired domain data. Additionally, we introduce precision@d, recall@d, and f1-score@d metrics, which measure the overlap rate of centerline coordinate sets to assess the similarity of multi-line segments. Our method extracts high-quality centerline regions, achieving performance comparable to fully-supervised training.