肝癌是台灣最常見的惡性腫瘤。在臨床上，肝癌的狀態可分為四個階段，階段越晚，病情越嚴重，而發現或病情越晚，死亡率越高。因此，判斷肝癌惡性腫瘤病理期別是個很重要的議題。

傳統上，MRI圖像可以幫助醫生直接確定第三和第四期。但是，對於第一階段和第二階段，應進行肝臟切片和手術分析，以鑑定階段。這種手術會給患者帶來痛苦和副作用。因此，在這項工作中，我們計劃開發一種實用的工具，該工具採用機器學習方法來結合患者數據和MRI來預測肝癌的分期。如果該工具的準確性很高，則可以幫助醫生加快判斷速度並節省醫療資源。對於患者來說，也可以避免侵入性手術。

我們的框架主要可以分為四個部分： (1) 患者數值資料、(2) 患者影像資料、(3)患者病理標籤特徵、(4) 預測模型。前三個部分是不同類型的醫學數據，將分別對其進行處理以進行分析和預測。最後一部分整合了前三個部分，並最終預測了肝癌的病理分期。 (pT1或pT2)

在第一部分中，(1) 患者數值資料。我們分析患者的個人數據，包括基本數據和血液採樣值，最後使用監督方法SVM進行預測。在第二部分中 (2) 患者影像資料，我們有九種不同的MRI圖像；我們獨立訓練這些圖像。因此，我們得到了9個CNN模型，並且每個CNN模型都將具有預測結果。第三部分是 (3)患者病理標籤特徵。通過MR圖像進行特徵和病理分析，最後選擇與該時期相關的四個特徵。第四部分是 (4) 預測模型。在最後一部分中，我們將使用機器學習模型來整合前三個部分。使用的功能包括第一部分和第二部分的預測(總共10個預測) 和第三部分的四個特徵。通過預測模型進行整合後，我們的工作最終將做出預測。

在我們的實驗中，肝癌病理期別( pT1或pT2）的預測準確性可達到91.28% (精度= 93.02％，召回率= 89.89％)。我們得出的結論是，我們的模型將幫助醫生判斷他們的差異，並且將是一個很好的臨床應用。

Liver cancer is the most common malignant tumor in Taiwan. Liver cancer is divided into four stages, the bigger the number, the more serious the condition is, and the later detection or serious condition, the higher the mortality rate. Therefore, it is an important issue to judge the stage of malignant tumor progression of liver cancer.

Traditionally, MRI images can help doctors directly determine the third and fourth phases. In the first and second phases, the liver should be sliced and analyzed to identify them. Therefore, in this work, I plan to use patient data and MRI to address a prediction task. The combination of images can avoid the operation to judge the first phase, which can help doctors speed up judgment and save medical resources. For patients, it can also avoid invasive surgery.

In our work, mainly can be divided into four parts. (1) Patient personal data, (2) Patient image data, (3) Patient pathological labels, (4) Predictive model. The first three parts are three different types of data, which are processed separately for analysis and prediction. The last part integrates the first three parts and makes the final prediction of the pathological stage of liver cancer. (pT1 or pT2)

In the first part, (1) patient personal data. We analyze the value of the patient's personal data, including the basic data and blood sampling values, and finally use the supervised method: SVM to make a prediction. In second part, (2) patient image data, we have nine different kinds of MR Image; we train these images independently. Therefore, we get nine CNN models, and each CNN model will have a predicted result. The third part is (3) patient pathological labels. Through MR Images, feature and pathological analysis were performed, and finally four features related to the period were selected. The fourth part is (4) predictive model. In the last part, we will integrate the first three parts using machine learning models. Used features are included the predictions for the first and second parts (10 predictions in total) and four features in the third part. Our work will eventually make a prediction after integration through the predictive model.

In our experiments, the prediction accuracy of the pathological stage (pT1 or pT2) can be achieved at 91.28% (Precision=93.02%, Recall=89.89%). Our model will help doctors judge their differences and will be a good clinical application.

[1] Alex Krizhevsky, Ilya Stuskever, Geoffrey E.Hinton. ImageNet classification with deep convolutional neural networks. NIPS'12 Proceedings of the 25th International Conference on Neural Information Processing Systems - Volume 1,2015.
[2] Elaheh Moradi, Antonietta Pepe, Christian Gaser, Heikki Huttunen, Jussi Tohka. Machine learning framework for early MRI-based Alzheimer's conversion prediction in MCI subjects. Author links open overlay panel. NeuroImage. Neuroscience Method, 2015.
[3] G. RätschT. OnodaK.-R. Müller. Soft Margins for AdaBoost. Kluwer Academic, 2001.
[4] Hong-Wei Ng, Viet Dung Nguyen, Vassilios Vonikakis, Stefan Winkler. Deep Learning for Emotion Recognition on Small Datasets using Transfer Learning. ICMI '15 Proceedings of the 2015 ACM on International Conference on Multimodal Interaction, 2015.
[5] Michael CollinsRobert E. SchapireYoram Singer. Logistic Regression, AdaBoost and Bregman Distances. Kluwer Academic, 2002.
[6] Patrick Ferdinand Christ, Mohamed Ezzeldin A. Elshaer, Florian Ettlinger, Sunil Tatavarty, Marc Bickel, Patrick Bilic, Markus Rempfler, Marco Armbruster, Felix Hofmann, Melvin D’Anastasi, Wieland H. Sommer, Seyed-Ahmad, AhmadiBjoern H. Menze. Automatic Liver and Lesion Segmentation in CT Using Cascaded Fully Convolutional Neural Networks and 3D Conditional Random Fields. International Conference on Medical Image Computing and Computer-Assisted Intervention 2016.
[7] R. Vivanti MSc, A. Ephrat1 MSc, L. Joskowicz PhD, O.A. karaaslan, N. Lev-Cohain2 MD, J. Sosna MD. Automatic liver tumor segmentation in follow-up CT studies using Convolutional Neural Networks. Prediction of Infant MRI Appearance and Anatomical Structure Evolution Using Sparse Patch-Based Metamorphosis Learning Framework, 2015.
[8] Sajid Iqbal, M. Usman Ghani, Tanzila Saba, Amjad Rehman. Brain tumor segmentation in multi‐spectral MRI using convolutional neural networks (CNN). Microscopy Research and Technique, 2018
[9] Salvatore C, Cerasa A, Castiglioni I, Gallivanone F, Augimeri A, Lopez M, Arabia G, Morelli M7 Gilardi MC, Quattrone A. Machine learning on brain MRI data for differential diagnosis of Parkinson's disease and Progressive Supranuclear Palsy. Journal of Neuroscience Methods, 2014.
[10] Sérgio Pereira, Adriano Pinto, Victor Alves, Carlos A. Silva. Brain Tumor Segmentation Using Convolutional Neural Networks in MRI Images. IEEE, 2013.
[11] Shuying Liu, Weihong Deng. Very deep convolutional neural network-based image classification using small training sample size. IEEE, 2015.
[12] Tianqi Chen, Carlos Guestrin. XGBoost: A Scalable Tree Boosting System. KDD '16 Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2016.
[13] Y. Lecun, L. Bottou, Y. Bengio, P. Haffner. Gradient-based learning applied to document recognition. IEEE, 1998.
[14] Yi-Chen Chou, et.al. Gadoxetic acid-enhanced magnetic resonance imaging can predict the pathologic stage of solitary hepatocellular carcinoma. World Journal of Gastroenterology, 2015.
[15] Zacharaki EI, Wang S, Chawla S, Soo Yoo D, Wolf R, Melhem ER, Davatzikos C. Classification of brain tumor type and grade using MRI texture and shape in a machine learning scheme. Magnetic Resonance in Medicine, 2009.
[16] Zhou LQ, Wang JY, Yu SY, Wu GG, Wei Q, Deng YB, Wu XL, Cui XW, Dietrich CF. Artificial intelligence in medical imaging of the liver. World Journal of Gastroenterology, 2015.