| 摘要: | 前庭神經鞘瘤是一種罕見且具有挑戰性的神經外科疾病，主要生長在內耳前庭神經區域，對患者生活有深遠影響。現今主要依賴核磁共振（MRI）進行診斷，但在病灶評估和治療策略制定方面仍面臨挑戰。常見治療方式包括手術和伽馬刀治療，需根據患者健康狀況和腫瘤特性制定適當計劃。
本研究使用前庭神經鞘瘤患者的T1和T2核磁共振加權影像合併伽馬刀治療表格資料，利用深度學習模型進行治療效果預測。我們收集了雙和醫院107位患者的治療前核磁共振影像，根據治療效果分為有效組和無效組，並整合每位患者的性別、年齡、腫瘤位置及大小、以及治療計畫的相關資料。
研究步驟包括影像預處理和臨床治療數據合併。首先，將核磁共振影像裁切為統一尺寸並進行正規化處理，以確保模型能夠有效學習影像中的關鍵特徵。接著，使用深度學習模型處理患者的T1和T2影像，並將其與表格數據結合，生成訓練集和測試集。模型結構包括輸入層、InceptionV3模型層、卷積和池化層、展平和Dropout層、數據融合層、全連接層和輸出層。並使用Adam優化器，分類交叉熵為損失函數，並監控分類準確度和AUC指標。
2
在模型訓練過程中使用K-fold交叉驗證，以確保模型的穩定性和可靠性。經過訓練，我們的模型在使用T1與T2影像結合伽馬刀治療相關表格資料進行預測後，達到準確度0.82，精準度0.84，召回率0.82，F-1分數0.83，AUC=0.76。結果顯示，結合多模態影像數據和表格數據的深度學習模型能有效預測加馬刀治療效果，協助臨床醫師制定個別性的治療計劃。
本研究展示了深度學習在醫學應用中的潛力，為未來深入研究和臨床實踐提供基礎，有望提高前庭神經鞘瘤的診斷和治療精確性，為患者提供優質醫療服務。
關鍵字：前庭神經鞘瘤、深度學習模型、核磁共振影像、伽馬刀治療
Vestibular schwannoma is a rare and challenging neurosurgical condition that primarily grows in the vestibular nerve area of the inner ear, significantly impacting patients' lives. Diagnosis currently relies heavily on magnetic resonance imaging (MRI), but challenges remain in lesion assessment and treatment strategy formulation. Common treatment options include surgery and Gamma Knife treatment, with appropriate plans tailored based on the patient's health status and tumor characteristics.
This study aimed to predict Gamma Knife treatment outcomes in Vestibular Schwannoma patients using deep learning models by combining MRI images and clinical data. Specifically, we utilized pre-treatment T1-weighted and T2-weighted MRI images from 107 patients at Shuang Ho Hospital. Patients were categorized into effective and ineffective groups based on their treatment outcomes. The clinical data in this study included each patient's gender, age, tumor location, tumor size, and specific details of the treatment plan.
The research steps included image preprocessing and data integration. First, the MRI images were cropped to a uniform size and normalized to ensure the model effectively learns key features from the images. Next, a deep learning model processed the patients' T1-weighted and T2-weighted images, combining them with the tabular data to create training and test sets. The model structure included convolutional layers, max pooling layers, fully connected
4
layers, and output layers. The Adam optimizer was used, with categorical cross-entropy as the loss function, and accuracy and AUC metrics were monitored.
We used K-fold cross-validation during model training to ensure robustness and reliability. After training, our model achieved an accuracy of 0.82, a precision of 0.84, a recall of 0.82, an F-1 score of 0.83, and an AUC of 0.76 when predicting treatment outcomes by combining T1-weighted and T2-weighted images with Gamma Knife treatment-related tabular data. The results indicate that a deep learning model combining multimodal imaging and tabular data can effectively predict Gamma Knife treatment outcomes, assisting clinicians in formulating personalized treatment plans.
This study demonstrates the potential of deep learning in medical applications, providing a foundation for further research and clinical practice. It promises to improve the accuracy of vestibular schwannoma diagnosis and treatment, offering high-quality medical services to patients.
Keywords: Vestibular schwannoma, deep learning model, magnetic resonance imaging, Gamma Knife treatment |
