| 摘要: | 背景動機
腦室周圍白質腦病變 (Periventricular leukoencephalopathy, PLV) 已經成為當前醫學研究的重要議題，特別是在神經系統疾病的診斷和監測中。對於臺北榮民總醫院癌病中心 (Taipei Veterans General Hospital Cancer Center)所接觸的腦部轉移性腦瘤患者而言，對這些腦部白化區域的詳細和準確分析顯得尤為重要。然而，傳統的影像分析方法可能限於主觀評估，容易產生偏差。這並不能完全滿足臨床上的需求。因此，我們認為使用先進的人工智慧技術，特別是深度學習，可能會為這一領域帶來突破。
研究目的
本研究旨在利用 CNN InceptionV3、ResNet101 和 VGG19 深度卷積神經網路模型，對臺北榮民總醫院癌病中心的腦部核磁共振(Magneticresonance imaging ,MRI)影像數據進行分析，進而計算腦部實質白化區域的嚴重程度。通過這種方式，我們希望提供一種更高效、準確的方法，來協助臨床醫師進行疾病評估。
研究方法
我們首先收集了臺北榮民總醫院癌病中心患者的腦部核磁共振影像數據。接著，利用深度學習技術，尤其是深度卷積神經網路模型，對這些影像數據進行訓練和優化。我們特別選擇了三個預先訓練的 CNN 模型 InceptionV3、ResNet101 和 VGG19，並使用遷移學習方法進行微調。我們的模型設定將腦部白化程度分為 Grade 0 至3。在模型訓練過程中，我們對不同的參數組合進行了多次嘗試，以確保獲得最佳效果。考慮到模型經訓練後，應用到新數據並作出準確預測能力的重要性，我們採用五折交叉驗證的方法確保結果的穩定性和可靠性。此外，我們也使用了多種效能指標，如kappa 係數、準確率和 F1 score、weighted precision 等，以全面評估模型的
性能。
研究結果
雖然我們的研究已有初步成績，但仍面臨一些挑戰和限制。我們正持續進行模型的優化工作，以期待未來能夠減少對人工標註的依賴。初步通過我們在人工標註輔助下的深度學習模型，我們可以自動且準確地分類腦部實質白化區域的嚴重程度。實驗結果顯示，三種模型均展現高準確性，其中，InceptionV3 模型在訓練過程中達到了100%的準確率，ResNet101 和VGG19模型也展現出高達97.5%和95%的準確性。
結論
人工智慧利用深度學習技術在腦室周圍白質腦病變在手工標註分析中展現出了其強大的潛力。我們可以更精確的分析核磁共振或電腦斷層(Computed Tomography ,CT)掃描圖像，以辨識和預測白質腦病的早期發展。我們期望這些新的分析方法能夠在未來臨床實踐中為臨床醫生提供更加精確、客觀和迅速的評估工具，從而改善病人的治療和照護品質。
Background
Periventricular leukoencephalopathy (PVL) has emerged as a significant focus of current medical research, particularly in the diagnosis and monitoring of neurological disorders. For patients with brain metastases treated at the Taipei Veterans General Hospital Cancer Center, a detailed and accurate analysis of these periventricular white matter lesions is of paramount importance. However, traditional image analysis methods may be limited by subjective assessments, leading to potential bias and falling short of clinical requirements. Therefore, we believe that the application of advanced artificial intelligence techniques, particularly deep learning, holds the potential for breakthroughs in this field.
Research purposes
This study aims to utilize CNN models, specifically InceptionV3, ResNet101, and VGG19 deep convolutional neural networks, to analyze brain MRI data from Taipei Veterans General Hospital Cancer Center and calculate the severity of periventricular leukoencephalopathy. Through this approach, we intend to provide a more efficient and accurate method to assist clinicians in disease assessment.
Research methods
We initially collected brain MRI data from patients at the Taipei Veterans General Hospital Cancer Center. Subsequently, using deep learning techniques, especially deep convolutional neural network models, we trained and fine-tuned these image data. We specifically chose three pre-trained CNN models: InceptionV3, ResNet101, and VGG19, and employed transfer learning for fine-tuning. Our model categorizes the severity of white matter lesions into Grade 0 to 3. During model training, we experimented with various parameter configurations to ensure optimal performance. Considering the importance of the model's ability to make accurate predictions on new data after training, we employed five-fold cross-validation to ensure the stability and reliability of the results. Additionally, we used multiple performance metrics, such as kappa coefficient, accuracy, F1 score, weighted precision, etc., to comprehensively evaluate the model's performance.
Result
Although our research has yielded preliminary results, we still face certain challenges and limitations. We are actively engaged in ongoing model optimization efforts, with the aim of reducing our reliance on manual annotations in the future. Through our deep learning models assisted by manual annotations, we have achieved the capability to automatically and accurately classify the severity of periventricular leukoencephalopathy in the brain regions. Experimental results have demonstrated high levels of accuracy across all three models. In particular, the InceptionV3 model reached 100% accuracy during the training process, while the ResNet101 and VGG19 models achieved accuracy levels of up to 97.5% and 95%, respectively.
Conclusion
Artificial intelligence, utilizing deep learning techniques, has demonstrated its significant potential in the analysis of periventricular leukoencephalopathy through manual annotations. This allows for a more precise analysis of magnetic resonance imaging (MRI) or computed tomography (CT) scan images, aiding in the identification and early prediction of white matter brain diseases. We anticipate that these novel analytical methods will provide clinical practitioners with more accurate, objective, and expeditious assessment tools in future clinical practice, thereby enhancing the quality of patient care and treatment. |
