基於機器學習的純音聽力測試聽力圖解譯

Abstract

根據世界衛生組織統計，在2050年，世界將有1/4的人口經歷聽力損失問題。無論造成聽力損失的原因為何，我們都需要進行聽力測試以記錄聽力損失程度、追蹤是否惡化，並給予適當的治療。 聽力圖的記錄格式對人類來說是方便閱讀的，但隨著機器學習、深度學習成為流行且有力的研究方法，若我們想要以機器來研究聽力問題，我們就需要將聽力圖由人類可讀的圖表形式，轉換為機器可讀的數字化形式。 此外，聽力圖數位化也能輔佐臨床工作者。在進行聽力檢查時，有許多不同的參數需要考慮，包括嚴重程度、對稱性、圖形形狀、聽損種類等。這些參數有著多種不同的定義，並且可能隨時進行調整。聽力圖數位化在臨床上有著幾個優勢，包括具有可核對的特性，能夠依據地區情況進行因地制宜的分析，並且相較圖片形式更能夠節省儲存空間。聽力圖數位化有助於協助臨床工作者維持其現有的作業流程，同時也能夠贏得他們的信任。然而，需要明確指出的是，這種方法並非旨在取代臨床工作者的判讀工作，而是成為其最佳的輔助角色，協助他們更準確地進行診斷工作。 本論文提出一個多階段的模型，結合YOLOv5與光學字元辨識，達到迅速且準確率達88%的端到端聽力圖數位化結果，希冀能成為日後聽力學相能研究的助力。

According to the World Health Organization statistics, by 2050, one in four people worldwide will experience hearing loss. Regardless of the cause of hearing loss, it is important to conduct hearing tests to assess the degree of hearing loss, monitor any deterioration, and provide appropriate treatment. While the recording format of audiograms is convenient for human readability, with machine learning and deep learning becoming popular and beneficial research approaches, if we want to study hearing issues using machines, we need to transform audiograms from human-readable chart formats into machine-readable digital formats. Furthermore, audiogram digitization can provide valuable assistance to clinical practitioners. When conducting auditory assessments, numerous parameters must be considered, including severity, symmetry, graphical shape, types of hearing loss, and more. These parameters come with various definitions that may be subject to adjustments over time. Audiogram digitization offers several advantages within the clinical realm. It possesses an auditability feature, allowing tailored analyses based on regional conditions, and notably, it is more space-efficient compared to image formats. This digitization process contributes to aiding clinical professionals in maintaining their existing workflow while also earning their trust. However, it is important to emphasize that this approach is not intended to replace the interpretative tasks of clinical professionals. Instead, it is designed to serve as an optimal supplementary role, assisting them in conducting diagnoses with enhanced accuracy. In this paper, we propose a multi-stage model that combines YOLOv5 and optical character recognition(OCR) to achieve rapid and accurate end-to-end digitization of audiograms, with an accuracy rate of 88%. We hope that this model can serve as a valuable tool for future research in audiology.