研究背景：聽力損失為全球重要公共衛生議題，其中慢性中耳炎所造成之鼓膜穿孔為臨床常見且可治療之病因。鼓膜穿孔會破壞外耳與中耳之聲壓傳遞機制，造成傳導性聽力損失，臨床上多以氣骨導差（air-bone gap， ABG）量化。純音聽力檢查雖為評估聽力之標準方法，但其執行仰賴病人配合、專業設備與適當檢查環境；在幼童、高齡者、認知功能受損者、基層醫療或資源受限場域中，常無法即時取得可靠聽力資料。相較之下，耳鏡與耳內視鏡影像為臨床例行取得之資料，能直接呈現鼓膜穿孔之大小、位置與形態，具有轉化為功能性評估資料來源之潛力。 研究目的：本論文旨在建立一套以深度學習為基礎之鼓膜穿孔影像分析架構，將耳鏡影像中可觀察之鼓膜與穿孔形態轉化為可量化且可重現之幾何特徵，並進一步應用於兩個臨床問題：其一為術前傳導性聽力損失之 ABG 評估，其二為第一型鼓膜成形術後聽力預後之預測。透過此一架構，本研究嘗試將傳統以質性判讀為主之鼓膜影像，延伸為可支援聽力功能評估與臨床決策輔助之量化資料來源。 研究方法：本研究採回溯性設計，整體資料包括三個主要部分。第一，納入 1，164 張門診鼓膜內視鏡影像，包括 1，014 張完整鼓膜影像與 150 張鼓膜穿孔影像，用於建立 Mask R-CNN 深度學習分割模型，以自動分割鼓膜與穿孔區域。第二，納入具有鼓膜穿孔與聽力資料之 225 張影像，包括 150 張門診穿孔影像與 75 張手術室穿孔影像，用於萃取穿孔面積比例、穿孔質心距離、穿孔至聽小骨參考點相對距離及穿孔角度等幾何特徵，並建立術前 ABG 預測模型。第三，納入 2020 至 2024 年間接受經耳道內視鏡第一型鼓膜成形術且術後達解剖成功之單純性鼓膜穿孔病人 121 例，結合年齡、術前 ABG 與術前影像幾何特徵，建立 ABG closure 與術後 ABG 之預後預測模型。模型表現分別以分割效能指標、相關性分析、迴歸模型效能、內部驗證及獨立測試資料進行評估。 研究結果：深度學習分割模型於鼓膜與穿孔區域分割皆展現良好表現，並可產生與人工標註相近之影像幾何特徵。術前 ABG 分析顯示，穿孔面積比例為影響傳導性聽力損失之主要影像因子，與 ABG 呈正相關；位置相關特徵則提供次要但具解釋意義之空間資訊。在術前 ABG 預測方面，二次迴歸模型較單純理論模型具有較佳表現，顯示影像幾何特徵可有效連結鼓膜穿孔形態與傳導性聽力損失。在術後聽力預後模型中，Elastic Net 模型於獨立測試資料集中預測 ABG closure 之 R² 為 0.418、RMSE 為 4.39 dB；使用 Mask R-CNN 自動特徵時，R² 為 0.421、RMSE 為 4.38 dB。術後 ABG 預測之 R² 為 0.363、RMSE 為 4.36 dB；使用自動特徵時，R² 為 0.366、RMSE 為 4.35 dB，顯示自動化影像特徵可達與人工標註相近之預測表現。 結論：本論文建立一套由鼓膜影像取得、深度學習自動分割、幾何特徵萃取、術前 ABG 預測至術後聽力預後評估之整合性影像分析架構。研究結果顯示，鼓膜穿孔影像中所蘊含之幾何資訊，尤其是穿孔面積比例，與傳導性聽力損失具有明確關聯；結合術前聽力資料與影像幾何特徵後，亦可合理預測第一型鼓膜成形術後之聽力改善與殘餘 ABG。本研究證實耳鏡影像可由傳統質性判讀工具，進一步轉化為可量化、可重現且具臨床決策輔助潛力之功能性資料來源，並為人工智慧於耳科聽力功能評估與手術預後預測之轉譯應用提供實證基礎。;Background: Hearing loss is a major global public health issue. Tympanic membrane perforation caused by chronic otitis media is a common and treatable cause of conductive hearing loss. Tympanic membrane perforation disrupts the sound pressure transmission mechanism between the external and middle ear， leading to conductive hearing loss， which is commonly quantified using the air-bone gap (ABG). Although pure-tone audiometry is the standard method for hearing assessment， its performance depends on patient cooperation， specialized equipment， and an appropriate testing environment. In children， older adults， patients with cognitive impairment， primary care settings， or resource-limited environments， reliable audiometric data may not always be immediately available. In contrast， otoscopic and otoendoscopic images are routinely obtained in clinical practice and directly demonstrate the size， location， and morphology of tympanic membrane perforations， suggesting their potential as a functional source of information for hearing assessment. Objective: This dissertation aimed to develop a deep learning–based image analysis framework for tympanic membrane perforations. The framework was designed to transform otoscopic image findings into quantifiable and reproducible geometric features and to apply these features to two clinically relevant tasks: preoperative assessment of conductive hearing loss using ABG prediction and prediction of postoperative hearing outcomes after type I tympanoplasty. Through this integrated framework， this dissertation sought to extend otoscopic images from qualitative visual interpretation to quantitative data that may support functional hearing assessment and clinical decision-making. Methods: This retrospective study consisted of three main datasets and analytical components. First， 1，164 outpatient otoendoscopic images， including 1，014 intact tympanic membrane images and 150 perforated tympanic membrane images， were used to develop a Mask R-CNN deep learning segmentation model for automatic segmentation of the tympanic membrane and perforation regions. Second， 225 perforation images with corresponding audiometric data， including 150 outpatient perforation images and 75 intraoperative perforation images， were used to extract geometric features， including perforation area ratio， perforation centroid distance， relative distance from the perforation to the ossicular reference point， and perforation angle， for preoperative ABG prediction. Third， 121 patients with simple tympanic membrane perforations who underwent transcanal endoscopic type I tympanoplasty between 2020 and 2024 and achieved anatomical success were included for postoperative prognostic modeling. Age， preoperative ABG， and preoperative image-derived geometric features were integrated to predict ABG closure and postoperative ABG. Model performance was evaluated using segmentation metrics， correlation analyses， regression model performance， internal validation， and independent test datasets. Results: The deep learning segmentation model demonstrated good performance in segmenting both the tympanic membrane and perforation regions， and the automated image-derived geometric features were comparable to those obtained from manual annotations. In the preoperative ABG analysis， perforation area ratio was the primary image-derived factor associated with conductive hearing loss and showed a positive correlation with ABG， whereas location-related features provided secondary but interpretable spatial information. For preoperative ABG prediction， the quadratic regression model outperformed the theoretical model， indicating that image-derived geometric features can effectively link tympanic membrane morphology with conductive hearing loss. In the postoperative prognostic model， the Elastic Net model achieved an R² of 0.418 and an RMSE of 4.39 dB for predicting ABG closure in the independent test dataset. When Mask R-CNN–derived automated features were used， the R² was 0.421 and the RMSE was 4.38 dB. For postoperative ABG prediction， the Elastic Net model achieved an R² of 0.363 and an RMSE of 4.36 dB， whereas the model using automated features achieved an R² of 0.366 and an RMSE of 4.35 dB. These findings indicate that automated image-derived features can achieve predictive performance comparable to manual annotations. Conclusions: This dissertation established an integrated image analysis framework encompassing tympanic membrane image acquisition， deep learning–based automated segmentation， geometric feature extraction， preoperative ABG prediction， and postoperative hearing outcome assessment. The results demonstrate that geometric information embedded in tympanic membrane perforation images， particularly perforation area ratio， is clearly associated with conductive hearing loss. When combined with preoperative audiometric data， image-derived geometric features can also reasonably predict hearing improvement and residual ABG after type I tympanoplasty. This study supports the transformation of otoscopic images from conventional qualitative interpretation into quantifiable， reproducible， and clinically meaningful functional data with potential value for decision support， and provides empirical evidence for the translational application of artificial intelligence in otologic functional assessment and surgical prognostication.
