思覺失調症 (Schizophrenia, SZ) 是心智的異常狀態，特徵為無法區分何者為真實，主要病徵包含妄想以及幻覺。我們使用病患的功能性磁振造影 (functional Magnetic Resonance Imaging, fMRI)，藉由自動解剖標籤 (Automated Anatomical Labeling) 將大腦分區，計算每個區域之間的相關係數，形成功能性連結矩陣 (Functional Connectivity Matrix)，將不同分區的特徵用在不同的模型進行判別，作為精神科醫師的決策支援輔助，藉以提升臨床診斷準確率。 研究的實驗資料來自臺北榮民總醫院精神科，包含220筆思覺失調症患者及220筆健康者(Healthy Control, HC) 的功能性磁振造影，性別及年齡比例相似非常接近。我們使用功能性連接矩陣探討不同種大腦分區方式、不同機器學習分類器、不同年齡層、不同性別間判別的效果。在大腦分區方式以AAL-3為最優秀，分類器則是使用支持向量機最好，準確率可達85.50%，而不同年齡及性別所比較出來的結果也顯示，同性質的資料所訓練的分類器能夠有較好的判別效果。 我們為求達到更好的判別效果，希望能以更多的資料來訓練模型，但是醫療相關影像的取得較難，成本也較高，因此希望能在有限的標記資料下能夠擴增成更大的資料集。功能性連接矩陣並不像一般影像資料集，能經由傳統的資料擴增方式如旋轉、鏡像或平移等進行擴充，因此我們希望能藉由生成對抗網路自訓練集的學習，生成更多樣本以訓練模型讓其能涵蓋不同的資料分布，提升判別的效果，在十次的實驗中訓練完生成對抗網路後，以擴增後的資料集進行訓練，羅吉斯回歸分類器的準確率能自84.77% 提升到87.05%，支援向量機的準確率能自85.23% 提升到86.36%。 ;Schizophrenia is an abnormal state of the brain and mind with main symptoms including delusions and hallucinations so that fail to distinguish the real and virtual worlds. We use functional Magnetic Resonance Imaging (fMRI) to partition the brain by automatic anatomical labeling, calculate the correlation coefficient for each pair of ROIs, and form a functional connectivity matrix. The features of different parcellations are used in different models for Schizophrenia discrimination. The experimental data came from Taipei Veterans General Hospital, Taiwan. A total of 440 resting fMRIs were originated from 220 patients diagnosed with schizophrenia and 220 patients as healthy controls, in which the gender, and age distributions were very close. We use the functional connectivity matrix to explore the effects of different brain parcellation methods, different machine-learning classifiers, different age and gender groups. We find that using a support vector machine model with AAL-3 achieves the best accuracy score of 85.50%. The results also reveal that different classification models trained on the same ages and genders group had better performance. In order to achieve a better discrimination performance, we augment the dataset with more labeled data using a conditional Generative Adversarial Network (cGAN). By the cGAN-based data argumentation, the accuracy of the logistic regression classifier can be improved from 84.77% to 87.05%, and the support vector machine can be enhanced from 85.23% to 86.36%.
