在醫學影像中，前縱膈腔的腫瘤看起來非常相似，在影像上目前還沒有一個標準流程去判斷胸腺瘤或是淋巴瘤，只能憑藉著醫生的經驗。我們的目標是找到能定量分類前縱膈腔腫瘤的參數或是方法，來輔助醫生在臨床上診斷。在這份研究中，我們利用動態顯影磁振造影(Dynamic Contrast Enhanced Magnetic Resonance Imaging, DCE-MRI)這項醫學影像技術，並利用不同的生理模型去分析前縱膈腔腫瘤的血流灌注參數，再對這些參數使用機器學習的方法分類，來找出能夠分類前縱膈腔腫瘤的參數與分類方式，以建立處理動態顯影磁振造影的標準方法。 方法：本研究蒐集了50位來自成功大學附設醫院的病人DICOM資料；其中13位患有淋巴癌；37位患有胸腺瘤。我們自行開發程式圈選感興趣的區域(Volume of Interest, VOI)以及計算動脈輸入函數(Arterial Input Function, AIF)，並將其影像強度轉換成顯影劑在組織內的濃度。我們共使用三種不同的生理模型:Tofts Kermode(TK)、Extended Tofts Kermode(ETK)和Two-compartment exchange(2CXM)模型去分析灌注參數。再將灌注參數利用接受者操作特徵曲線(Receiver operating characteristic, ROC)以及三種不同的機器學習方法；決策樹(Decision tree)、支撐向量機(Support vector machine, SVM)和K近鄰演算法(K-nearest neighbor, KNN)進行分類，並觀察三種生理模型的滲透參數差異，以及比較不同的機器學習演算法的分類結果。 結果：我們利用了自己的程式透明化動態顯影磁振造影數據分析流程，並使用機器學習模型再分類胸腺瘤以及淋巴癌上，取得了80%的正確率。並且在分類胸腺瘤以及胸腺癌上取得正確率、特異度和敏感度三項都超過92%的判斷效果。 結論：可使用兩隔室模型分析動態顯影磁振造影數據，計算出灌注參數。利用灌注參數訓練而成的決策樹分類模型去分類前縱膈腔腫瘤，並再未來加上更多的前縱膈腔腫瘤資料加以優化模型，在臨床診斷時可提供可靠的建議。

In current clinical and imaging exams of anterior mediastinal tumors, one existing challenge exists in the differentiation of thymic tumors and lymphoma. Our aim is to
quantitatively classify anterior mediastinal tumor with image-derived parameters and classification models to assist doctors in clinical diagnosis. In this study, we used the dynamic contrast enhanced magnetic resonance imaging (DCE-MRI) as the main imaging modality. Moreover, we used three different compartment models to analyze the perfusion parameters from DCE-MRI data to evaluate the anterior mediastinal tumors. Method: In this study, we collected data from fifty patients, including imaging and clinical history, from the National Cheng Kung University Hospital. Biopsy or surgical pathology found 13 of them with lymphoma and 37 of them with the thymic tumor. We drew the volume of interest (VOI) and the arterial input function (AIF) with our in-house software, and converted the image intensity into the concentration of the contrast agent (CA) in the tissue. Moreover, we use three different compartment models, Tofts Kermode (TK) model, Extended Tofts Kermode (ETK) model, and two-compartment exchange (2CXM) model to calculate the perfusion parameters. We used the receiver operating characteristic (ROC) curve and three different machine learning methods including decision trees, support vector machine (SVM), and K-nearest neighbor (KNN). Finally, we evaluated the differences in the perfusion parameters of three compartment models and compared the classification results of different machine learning algorithms. Result: We make the process of DCE-MRI data analysis transparent with our in-house software. Moreover, we classified lymphomas and thymic tumors by our decision tree models and eventually reached 80% on accuracy. Furthermore, the accuracy, sensitivity, and specificity of the result of classifying thymomas and thymic carcinomas can all surpass 92%. Conclusion: Our software can analyze the DCE-MRI data with the two-compartments model and calculate the perfusion parameters. We used the perfusion parameters to train the decision tree classification model which can classify the anterior mediastinal tumors. We will add more tumor data in the future to optimize the classification model provide a reliable computer-assisted diagnosis in the clinical setting.

1. CANCER REGISTRY ANNUAL REPORT, 2017
TAIWAN, H.P.A.M.O.H.A.W. TAIWAN, Editor. 2019.
2. Engels, E.A. and R.M. Pfeiffer, Malignant thymoma in the United States: demographic patterns in incidence and associations with subsequent malignancies. Int J Cancer, 2003. 105(4): p. 546-51.
3. Thymoma and Thymic Carcinoma: Statistics. 2020; Available from: https://www.cancer.net/cancer-types/thymoma-and-thymic-carcinoma/statistics.
4. Ugalde, P.A., et al., Correlative anatomy for the mediastinum. Thorac Surg Clin, 2011. 21(2): p. 251-72, ix.
5. Whitten, C.R., et al., A diagnostic approach to mediastinal abnormalities. Radiographics, 2007. 27(3): p. 657-71.
6. Duwe, B.V., D.H. Sterman, and A.I. Musani, Tumors of the mediastinum. Chest, 2005. 128(4): p. 2893-909.
7. Wilkins, T.W.S.L.W., General Thoracic Surgery. 2009.
8. Nitz, W.R. and P. Reimer, Contrast mechanisms in MR imaging. Eur Radiol, 1999. 9(6): p. 1032-46.
9. Glover, G.H., Overview of functional magnetic resonance imaging. Neurosurg Clin N Am, 2011. 22(2): p. 133-9, vii.
10. Tofts, P.S., Modeling tracer kinetics in dynamic Gd-DTPA MR imaging. J Magn Reson Imaging, 1997. 7(1): p. 91-101.
11. Rohrer, M., et al., Comparison of Magnetic Properties of MRI Contrast Media Solutions at Different Magnetic Field Strengths. Investigative Radiology, 2005. 40(11): p. 715-724.
12. Tofts, P.S., et al., Estimating kinetic parameters from dynamic contrast-enhanced T(1)-weighted MRI of a diffusable tracer: standardized quantities and symbols. J Magn Reson Imaging, 1999. 10(3): p. 223-32.
13. Brix, G., et al., Microcirculation and microvasculature in breast tumors: pharmacokinetic analysis of dynamic MR image series. Magn Reson Med, 2004. 52(2): p. 420-9.
14. Sourbron, S., et al., Quantification of cerebral blood flow, cerebral blood volume, and blood-brain-barrier leakage with DCE-MRI. Magn Reson Med, 2009. 62(1): p. 205-17.
15. Tseng, C. Decision trees. 02 2017; Available from: https://chtseng.wordpress.com/2017/02/10/%E6%B1%BA%E7%AD%96%E6%A8%B9-decision-trees/?fbclid=IwAR2uAbXX5BJ8qjFQ2pSbELGIW5k4W1BghRc5ZegKCG9-cPWObusZVHWkpFA.
16. Romi, F., Thymoma in myasthenia gravis: from diagnosis to treatment. Autoimmune Dis, 2011. 2011: p. 474512.