| 摘要: | 吞嚥障礙指的是在吞嚥食物過程中遇到的困難。這些困難可能導致嗆咳、吸入性肺炎、脫水和營養不良等風險。早期發現吞嚥障礙並進行及時的吞嚥訓練可以有效幫助患者預防吸入性肺炎及相關併發症。因此，準確及時地診斷患者是否有吞嚥困難至關重要。
目前，診斷吞嚥困難的黃金標準評估工具是吞嚥攝影檢查（VFSS）。這種評估工具為臨床醫療專業人員提供了客觀且全面的診斷基礎，使其能夠精確觀察吞嚥過程中每個階段的生理結構和功能。然而，臨床專業人員需要投入大量的時間和人力來進行影像分析和報告撰寫。此外，影像分析結果的評分者間信度通常不理想。
近年來，研究人員開發了多種基於機器學習的VFSS自動分析方法。但較少有研究使用深度學習模型來檢測多個事件，並且沒有研究將模型預測的結果與現有的吞嚥障礙評分量表進行比較。
因此，我們的研究旨在使用深度學習模型（YOLOv9）來自動檢測多個與吞嚥障礙相關事件，並調查其與既有的吞嚥障礙評分量表（如MBSImp和DIGEST）的相關性或預測能力。結果顯示，我們訓練的YOLOv9模型達到了中等準確度，精度和召回率為70.06%，mAP_0.5為72.00%，顯示出其在臨床應用中的顯著潛力。相關性分析（F統計量=18.167，p值=0.002）表明，我們的模型能夠有效分類喉部食團殘留嚴重程度，且與MBSImp量表一致，這表明其具有增強臨床評估的潛力。我們相信，我們的模型可以幫助臨床醫療專業人員提高效率，並加強評估結果的一致性。此外，它還可以成為未來吞嚥評估工具開發的先驅。
Swallowing disorders refer to difficulties encountered during the process of swallowing food. These difficulties may lead to risks such as choking, aspiration pneumonia,dehydration, and malnutrition. Early detection of swallowing disorders and timely swallowing training can effectively help patients prevent aspiration, and related complications. Therefore, the accurate and timely diagnosis of whether a patient has swallowing difficulties is crucial.
Currently, the gold standard assessment tool for diagnosing swallowing difficulties is the Videofluoroscopic Swallow Study (VFSS). This assessment provides clinical healthcare professionals with an objective and comprehensive diagnostic basis, allowing precise observation of the physiological structures and functions at each stage of the swallowing process. However, clinical professionals need to invest a significant amount of time and manpower in image interpretation and report generation. Moreover, inter-rater reliability in the interpretation results is often suboptimal.
In recent years, researchers have developed several machine-learning based approaches for VFSS automatic analysis. However, few studies use deep learning models to detect multiple events, and no studies compare the models’ performances with existing dysphagia rating scales.
Therefore, our study aims to use a deep learning model(Yolov9) for automatically detecting multiple events and to investigate its correlation or predictive ability with established swallowing disorder rating scales such as MBSImp and DIGEST. The results show that our YOLOv9 trained model achieved a moderate level of accuracy, with a precision and recall of 70.06% and an mAP_0.5 of 72.00%, showcasing significant potential for clinical applications. The correlation analysis (F-statistic = 18.167, p-value = 0.002) indicates that our model effectively classifies pharyngeal residue severity. It aligns well with the MBSImP scale, suggesting its potential to enhance clinical evaluations. We believe that our model can assist clinical healthcare professionals in improving efficiency and enhancing the consistency of interpretation results. Additionally, it can serve as a pioneering approach for the development of future swallowing assessment tools. |
