| 摘要: | 聽神經瘤(Vestibular Schwannoma)為一種常見的良性腦部腫瘤，通常生長於負責傳遞聽覺訊息與平衡控制的前庭耳蝸神經上。隨著腫瘤體積增大，可能壓迫小腦、腦幹及周邊神經，導致眩暈、聽力損失、耳鳴與平衡障礙等症狀。目前臨床上常見的治療方式包括加馬刀放射治療與開顱手術，因此，精確掌握腫瘤在影像中的位置與體積對於治療決策至關重要。現行臨床上腫瘤的定位與量測多仰賴醫師以肉眼進行判讀，然而此方式不僅耗時，亦易受主觀判斷影響，缺乏客觀性與一致性。至今仍缺乏一套能有效輔助醫師準確判定腫瘤位置、體積及對小腦腦幹壓迫程度的自動化工具。
本研究提出一套基於深度學習的自動化分割與量化系統，針對聽神經瘤與小腦腦幹進行分析。我們採用注射對比顯影劑後T1權重影像(Contrast-enhanced T1-weighted Imaging, CE-T1WI)，並應用nnU-Net架構進行聽神經瘤及小腦腦幹的自動分割，進一步量化腫瘤對小腦腦幹的壓迫比例，並依據Koo′s 分級系統進行腫瘤嚴重程度之分類。本研究納入83位病患進行模型訓練與五折交叉驗證。在聽神經瘤的分割任務中，模型表現Dice Similarity Coefficient(DSC)為93.40±0.58%、精確率為94.27±1.20%、召回率為92.83±1.44%；而小腦腦幹分割的DSC為96.99±0.11%、精確率為96.96±0.24%與召回率為 97.03±0.14%。在分類模型部分，採用從579筆分割影像和空間標準化後分割影像中提取的特徵進行訓練與五折交叉驗證，四類(I–IV)分類準確率為82.38±1.94%、加權精確率為84.88±2.47%、加權召回率為82.38±1.94%、加權F1分數為82.83±1.80%，Kappa值為72.97±2.86%；二類(≤III vs. IV)分類準確率為91.53±2.22%、加權精確率為91.70±2.12%、加權召回率為91.53±2.22%、加權F1分數為91.52±2.23%，AUC高達97.60±1.22%。此外，本研究亦開發圖形化使用者介面，協助臨床醫師快速應用此系統於實務診療流程中。
本研究成果成功實現聽神經瘤及小腦腦幹結構之自動化分割與精確量化，提供醫師於臨床上判斷腫瘤生長情況、評估對周邊組織之壓迫程度，並協助監測腫瘤變化、術後殘留與復發。此系統具備高準確性與臨床實用性，有助於提升診斷效率與客觀性，進而優化治療決策品質。
;Vestibular schwannoma is a common benign brain tumor that typically arises from the vestibulocochlear nerve, which is responsible for transmitting auditory signals and maintaining balance. As the tumor grows, it may compress the cerebellum, brainstem, and surrounding nerves, leading to symptoms such as vertigo, hearing loss, tinnitus, and balance disorders. Current clinical treatments include Gamma Knife radiosurgery and craniotomy. Therefore, accurate localization and volumetric assessment of the tumor in medical images are critical for treatment planning. However, current clinical assessments rely heavily on manual interpretation by physicians, which is time-consuming and subject to inter-observer variability, lacking objectivity and consistency. To date, there remains a lack of automated tools that can assist clinicians in accurately identifying tumor location, measuring volume, and quantifying compression on the cerebellum and brainstem.
In this study, we propose an automated deep learning-based segmentation and quantification system for analyzing vestibular schwannoma and its impact on the cerebellum and brainstem. Contrast-enhanced T1-weighted imaging (CE-T1WI) was used as input, and the nnU-Net framework was adopted for automated segmentation of vestibular schwannomas and cerebellar-brainstem structures. The degree of compression exerted by the tumor on adjacent structures was quantitatively assessed, and tumor severity was classified based on the Koos grading system. A total of 83 patients were included for model training and five-fold cross-validation. For vestibular schwannoma segmentation, the model achieved a Dice Similarity Coefficient (DSC) of 93.40±0.58%, precision of 94.27±1.20%, and recall of 92.83±1.44%. For cerebellum and brainstem segmentation, the DSC reached 96.99±0.11%, precision 96.96±0.24%, and recall 97.03±0.14%.Regarding tumor severity classification, features were extracted from 579 segmented and spatially normalized images. The four-class (Koos I–IV) classification achieved an accuracy of 82.38±1.94%, weighted precision of 84.88±2.47%, weighted recall of 82.38±1.94%, weighted F1-score of 82.83±1.80%, and a Cohen’s Kappa of 72.97±2.86%. For binary classification (Koos ≤III vs. IV), the model achieved an accuracy of 91.53±2.22%, weighted precision of 91.70±2.12%, weighted recall of 91.53±2.22%, weighted F1-score of 91.52±2.23%, and an AUC of 97.60±1.22%. Additionally, a graphical user interface (GUI) was developed to facilitate the integration of this system into clinical workflows.
This study successfully demonstrates automated and precise segmentation and quantification of vestibular schwannomas and cerebellar-brainstem structures. The proposed system assists clinicians in evaluating tumor progression, assessing compression on adjacent structures, and monitoring tumor changes, postoperative residuals, or recurrence. With high accuracy and clinical applicability, this system enhances diagnostic efficiency and objectivity, ultimately supporting better-informed treatment decisions. |
