以人工智慧為基礎之兒童神經疾病腦部影像診斷模式－以妥瑞氏症為例

Abstract

腦影像檢查是兒童神經疾病診斷上非常重要的一項工具，但是並不是所有的兒童神經疾病都會有很明顯腦影像的變化，因此如何利用一些工具，善用影像的變化建立一個模式，來做一個判斷預測病人的預後，對臨床醫師的診斷以及對家長來說都很有幫助。在本研究中，我們選擇一個常見的兒童神經疾病¬－妥瑞氏症作為本研究的主題。在臨床上妥瑞氏症腦磁振造影影像不會有太大的差異。因此如何利用人工智慧建立新的診斷模式來評估妥瑞氏症病人的預後，是一個很大的挑戰。在本研究中，我們以腦磁振造影影像為分析資料，使用ROBEX去除非腦組織，然後透過AutoEncoder以及以3維卷積神經網路，並採用Fixed Offset的資料增強策略來建立妥瑞氏症預測模型，此模式可以達到最佳的準確性、敏感性和特異性。使用Fixed Offset的資料增強策略(每4個pixels取1個)，雖然整體敏感性下降，但準確性(63.96% vs.58.31%)和特異性(69.65% vs.51.90%)都大幅提升，對於嚴重的妥瑞氏症兒童的敏感性也提升(78.90% vs.76.30%)，證明此策略更能判斷妥瑞氏症兒童重症與輕症的腦結構是否更有一些差異。此外，我們的實驗結果也發現，資料增強策略的取樣頻率高低也會影響分類預測的效能，過高(每2個pixels取1個)或過低(每8個pixels取1個)的取樣頻率都會減低預測模型的效能。此外增加健康與病人的數目對模型的預測效能也有幫助，如果AutoEncoder和卷積神經網路使用更多受試者的影像資料來進行訓練，明顯的診斷為嚴重妥瑞氏症病人和健康受試者的準確性由70.42％提高到78.30％，而特異性也由70.55％提高到82.42％；雖然嚴重組跟輕症組的敏感性降低，尤其是輕症組的敏感性由49.28％降低到27.97％，不過這也反映出輕症的妥瑞氏症病童比較少會造成腦的影響或是健康受試者比較難和輕症的妥瑞氏症病人做區分，這跟我們的假說是吻合的。我們的研究也發現，嚴重性妥瑞氏症的病童通常比較有很顯著的腦異常，相對地較為輕症的妥瑞氏症病童腦異常的程度較為輕微或是沒有異常。這與我們先前的推測相吻合，這代表嚴重的妥瑞氏症比較會造成病童腦的異常，或是腦有異常的病童比較會造成嚴重的妥瑞氏症；相反的，輕症的妥瑞氏症病童通常沒有腦的異常或是腦的異常很輕微。我們將每一個妥瑞氏症病人的64組資料異常的部位組合以後，去檢視腦主要異常的地方和正常受試者的差異在哪裡。結果我們可以看到主要異常的位置在兩側基底核和前額葉的部位，當然有些個案也會影響到胼胝體後面的部位和腦白質，這與妥瑞氏症的致病機轉相吻合。由我們系列的實驗與研究發現，使用人工智慧用於妥瑞氏症病人腦影像異常的判斷是可行的。未來這種影像分析的模式也可以運用到其他的兒童神經疾病作為疾病對腦影響嚴重程度的判斷，當然也可以有商品化的可能性。

Brain imaging is a very important tool in the diagnosis of neurological diseases in children, but not all neurological diseases have obvious changes in brain imaging. Therefore, how to use some tools to establish a model to predict the prognosis of patients is very helpful for clinicians and parents. In the present study, we chose Tourette syndrome, a common neurological disorder in children, as the target of our research. Clinically, MRI of the brain in Tourette syndrome usually does not have prominent abnormality. Therefore, how to use artificial intelligence to establish a new diagnostic model to assess the prognosis of patients with Tourette syndrome becomes a big challenge. In the present study, we propose a deep-learning-based Tourette syndrome prediction model that uses ROBEX for removing non-brain tissue, and involves the use of AutoEncoder and 3D convolutional neural networks (3D CNN) to construct a classification model. We also develop a fixed- offset data augmentation strategy, generating multiple subsamples by selecting one pixel from every four pixels according a fixed offset from an original 3D brain image. Our experimental results found that this model architecture will reach the highest accuracy, sensitivity, and specificity of Tourette syndrome prediction based on brain MRI images. Compared with the random pick strategy, the use of fixed offset strategy for data augmentation improved the prediction accuracy (63.96% vs.58.31%), specificity (69.65% vs.51.90%), and the sensitivity of brain involvement in children with severe Tourette syndrome (78.90% vs. 76.30%), at the cost of the overall sensitivity. In addition, the sampling frequency will significantly affect the effectiveness of our proposed Tourette syndrome prediction model. Specifically, as compared to our proposed sampling frequency (selecting one pixel from every 4 pixels), higher (selecting one pixel from every 2 pixels) or lower (selecting one pixel from every 8 pixels) sampling frequency deteriorated the prediction effectiveness. Moreover, our experimental results also showed that increasing the number of subjects to train both AutoEncoder and 3D CNN increased the accuracy of the diagnosis of severe Tourette’s patients and healthy subjects from 70.42 % to 78.30%, and the specificity from 70.55% to 82.42%. Although the sensitivity of the severe group and the mild group was decreased, especially the sensitivity of the mild group, decreasing from 49.28% to 27.97%, this result suggested that children with mild Tourette syndrome would have fewer brain lesions or healthy subjects are more difficult to distinguish from mild Tourette’s patients, which is consistent with our hypothesis. Our study also found that children with severe Tourette syndrome usually have significant brain abnormalities, while children with milder Tourette syndrome have milder or no brain abnormalities. This is also consistent with our previous speculation, indicating that severe Tourette syndrome will cause more abnormalities in the brain of children, or children with brain abnormalities will more likely lead to severe Tourette syndrome; in contrast, children with mild Tourette syndrome usually have no or milder brain abnormalities. If we combined all 64 MRI subsamples, generated by our data augmentation strategy, for each Tourette patient to see where are the main locations of brain abnormalities differing from those in normal subjects. We can see that the main abnormalities are located in bilateral basal ganglion and frontal white matter. In some cases, Tourette syndrome also affects the splenium part of the corpus callosum and the white matter, which are consistent with the pathogenic mechanisms of Tourette syndrome. From our serial experiments, we have found that it is feasible to use artificial intelligence techniques to determine brain imaging abnormalities in patients with Tourette syndrome. In the future, our proposed model can also be applied to other neurological diseases in children to evaluate the severity of the impact of the disease on the brain. Of course, there is also the possibility of commercialization.