近年來全球人口結構逐漸高齡化，牙周疾病之盛行率亦隨之增加，牙周健康已成為高齡族群口腔照護中不可忽視的重要議題。牙周疾病若未能及早診斷與妥善追蹤，可能導致齒槽骨流失，進而影響咀嚼功能與整體生活品質，亦增加後續治療之醫療負擔。

在此背景下，如何在有限的醫療資源下提升診斷效率與治療品質，已成為當前醫療科技發展的重要課題。牙周囊袋深度為臨床評估牙周健康狀況與疾病嚴重程度的關鍵指標，其量測結果直接影響後續治療決策與療效追蹤。

目前臨床上常用之牙周囊袋量測方式以人工牙周探針為主，然而該方法高度仰賴操作者經驗與施力控制，易受探針角度等因素影響，導致量測結果存在主觀性與再現性不足的問題。此類限制使得人工量測在長期追蹤與跨操作者比較上，難以提供穩定且一致之量測依據。

因此，本研究自行開發超音波牙周囊袋深度量測系統，透過分析超音波回波訊號來量測牙周囊袋深度，藉由水作為超音波傳導介質，以達到維持相對穩定之量測條件，並建立較一致的量測條件，為了提升量測的準確性與再現性，本研究進一步結合小波轉換與深度學習模型，萃取回傳訊號中的特徵資訊，並建立牙周囊袋深度之分類機制，以進行自動化的深度判定。此方法可作為非接觸式量測，減少病患不適感，並提供一種具備客觀性、可重複性，且能降低人為誤差之牙周囊袋深度估測方式。

為評估所提出方法之分類性能與泛化能力，本研究設計兩組實驗。首先，在單一資料集實驗中，比較傳統訊號處理計算與卷積神經網路模型分類的表現，結果顯示傳統演算法不足以穩定對應深度，相較之下，卷積神經網路模型將三組量測數據以8:2 的比例切分為訓練集與測試集，以評估模型在相同量測條件下的分類能力。結果顯示模型在相同量測條件下可達到高度分類準確度。

其次，本研究進一步設計跨資料集驗證實驗，使用兩組資料進行訓練，並以另一組資料作為測試集，以評估模型在不同量測條件下的泛化能力。實驗結果顯示，當訓練資料與測試資料來自不同量測條件時，模型表現顯著下降，顯示超音波訊號特徵對量測條件具有高度敏感性。此結果指出，在實際應用中，維持一致的量測條件與標準化資料擷取流程，對於提升模型穩定性與泛化能力具有重要影響。

另一方面，豬隻牙周實際量測觀察顯示，水流強度控制與氣泡生成為影響超音波訊號穩定性的重要因素。當水流強度不足時，可能無法有效打開狹窄之牙齦縫；然而，過高的水流強度則可能產生氣泡與流場擾動，進而干擾超音波傳播並造成回波訊號不穩定。因此，未來超音波牙周囊袋深度量測系統之發展，除了提升模型在不同量測條件下的穩定表現外，亦需針對水流控制、氣泡抑制、耦合穩定性與即時訊號品質監測進行優化。

In recent years, as the global population continues to age, the prevalence of periodontal disease has increased, making periodontal health an increasingly important concern in oral care for older adults. If periodontal disease is not diagnosed early and properly monitored, it may lead to alveolar bone loss, thereby impairing masticatory function and overall quality of life. It may also increase the financial burden associated with subsequent treatment.

In this context, improving diagnostic efficiency and treatment quality within limited healthcare resources has become an important challenge in clinical dentistry and medical technology development.Periodontal pocket depth (PPD) is a key clinical indicator for assessing periodontal health and disease severity, and its measurement results directly influence subsequent treatment planning and the monitoring of therapeutic outcomes.

Currently, periodontal pocket depth in clinical practice is primarily measured using manual periodontal probes. However, this method is highly dependent on the operator's experience and probing force control, and it is easily affected by factors such as probe insertion angle. As a result, manual probing may lead to subjective measurement outcomes and limited reproducibility. These limitations make it difficult to obtain stable and consistent data for long-term follow-up and inter-operator comparisons.

To address these issues, this study developed an ultrasonic periodontal pocket depth measurement system that estimates periodontal pocket depth by analyzing ultrasonic echo signals. Water was used as the ultrasonic transmission medium to maintain relatively stable coupling and establish more consistent measurement conditions. To improve measurement accuracy and reproducibility, this study further integrated wavelet transform and deep learning models to extract features from the reflected ultrasonic signals and establish a depth classification mechanism for automated periodontal pocket depth determination. As a non-contact measurement technique, the proposed method may reduce patient discomfort while providing an objective and repeatable approach for estimating periodontal pocket depth.

To evaluate the classification performance and generalization ability of the proposed method, two sets of experiments were designed in this study. First, a single dataset experiment was conducted to compare the classification performance of traditional signal processing methods and a CNN-based model. The results showed that traditional algorithms were insufficient for stable depth estimation. In contrast, for the CNN-based approach, the three sets of measurement data were divided into training and testing sets at an 8:2 ratio to evaluate the model’s classification capability under identical measurement conditions. The results demonstrated that the CNN model achieved high classification accuracy under consistent measurement conditions.

Second, a cross-dataset validation experiment was further conducted, in which two datasets were used for training and the remaining dataset was used as the test set. This experiment was designed to evaluate the model's generalization ability under different measurement conditions. The results showed that when the model was evaluated across datasets obtained under different measurement conditions, its performance declined significantly. This finding indicates that ultrasound signal features are highly sensitive to variations in measurement conditions.

Overall, the findings of this study suggest that maintaining consistent measurement conditions and standardizing data acquisition procedures are crucial for improving model stability and generalization ability in practical applications. Practical observations from porcine periodontal measurements further indicated that water-flow control and bubble formation are important factors affecting ultrasonic signal stability. Therefore, future development of ultrasound-based periodontal pocket depth measurement should consider not only model robustness but also measurement-condition control, including water-flow optimization, bubble reduction, and signal quality monitoring.

[1] E. Bernabe, W. Marcenes, R. S. Abdulkader, L. G. Abreu, S. Afzal, F. N. Alhalaiqa,S. Al-Maweri, U. Alsharif, A. E. Anyasodor, A. Arora, and S. Asgary.Trends in the global, regional, and national burden of oral conditions from 1990 to 2021: A systematic analysis for the global burden of disease study 2021. The Lancet, 405(10482):897–910, 2025.
[2] Niklaus P. Lang and Jan Lindhe, editors. Clinical Periodontology and Implant Dentistry. John Wiley & Sons, 2015. 2 Volume Set.
[3] J. Ainamo and I. Bay. Problems and proposals for recording gingivitis and plaque.International Dental Journal, 25(4):229–235, 1975.
[4] D. D. Bosshardt. The periodontal pocket: Pathogenesis, histopathology and consequences.Periodontology 2000, 76(1):43–50, 2018.
[5] Saba Khan and Leyvee Lynn Cabanilla. Periodontal probing depth measurement:A review. Compendium of Continuing Education in Dentistry, 30(1):12–14, 16,18, 2009.
[6] M. A. Listgarten. Periodontal probing: What does it mean? Journal of Clinical Periodontology, 7(3):165–176, 1980.
[7] T. M. Hassell, M. A. Germann, and U. P. Saxer. Periodontal probing: Interinvestigator discrepancies and correlations between probing force and recorded depth. Helvetica Odontologica Acta, 17(1):38–42, 1973.57
[8] P. Birek, C. A. G. McCulloch, and V. Hardy. Gingival attachment level measurements with an automated periodontal probe. Journal of Clinical Periodontology, 14(8):472–477, 1987.
[9] U. Van der Velden and J. H. De Vries. Introduction of a new periodontal probe: The pressure probe. Journal of Clinical Periodontology, 5(3):188–197, 1978.
[10] Z. H. Hsieh, C. Y. Lai, N. H. Chen, and C. K. Yeh. Acoustic vortex-assisted thrombolysis treatment in a pulmonary embolism model using a miniature ultrasound catheter. Journal of Medical and Biological Engineering, 44(3):478–487, 2024.
[11] J. M. Thijssen. The history of ultrasound techniques in ophthalmology. Ultrasound in Medicine & Biology, 19(8):599–618, 1993.
[12] J. Hou, S. T. Rose, and M. K. Hinders. Ultrasonic periodontal probing based on the dynamic wavelet fingerprint. EURASIP Journal on Advances in Signal Processing, 2005(7):198348, 2005.
[13] M. Renaud, M. Gette, A. Delpierre, S. Calle, F. Levassort, F. Denis, and G. Y. Rochefort. Intraoral ultrasonography for the exploration of periodontal tissues: A technological leap for oral diagnosis. Diagnostics, 14(13):1335, 2024.
[14] Lei Fu et al. Posterior photoacoustic/ultrasound imaging of the periodontal pocket with a compact intraoral transducer. Photoacoustics, 28:100408, 2022.
[15] I. Daubechies. The wavelet transform, time-frequency localization and signal analysis. IEEE Transactions on Information Theory, 36(5):961–1005, 2002.