探討水腦症MRI影像於人工智慧分類之研究

The Classification of Hydrocephalus of MRI Using Artificial Intelligent Methods

姚依汶(Yi-Wen Yao)

指導教授 ： 陳泰賓

共同指導教授 ： 杜維昌(Wei-Chang Du)

義守大學/電機資訊學院/資訊工程學系/碩士(2019年)

摘要

臨床中使用磁振造影(Magnetic Resonance Imaging, MRI)診斷水腦症(Hydrocephalus)時，常使用Evan’s指數(Evan’s Index, EI)及蜘蛛網膜下腔不成比例擴大的腦積水(Disproportionately Enlarged Subarachnoid-space Hydrocephalus, DESH)作為定量指標。本研究探討人工智慧(Artificial Intelligence, AI)之分類方法，包含類神經網路(Artificial Neural Network)及卷積神經網路(Convolutional Neural Network, CNN)。
本研究採用回顧性分組實驗設計，根據臨床醫師診斷之報告共70例，包含31例正常案例及39例水腦症案例，共有48位男性(51±18.7歲)及22 位女性(49±22.8歲)，平均年齡為50.7歲。將案例進行影像分組後，正常影像與水腦症2D MRI影像分別為1073張與460張。本研包含了三種特徵萃取方法，包含5個一階統計量 (First-order Statistics)、20個灰階共生矩陣 (Gray-level Co-occurrence Matrix, GLCM)、70個Laws’紋理能量量測法(Laws’ Texture Energy Measures, LTEM)，以及280個結合GLCM及LTEM之特徵；經由T檢定找出40個具有顯著性(P-value<0.05)之特徵，透過40種特徵建立ANN水腦症分類模型。同時採用三種卷積神經網路模型進行分類水腦症，包含AlexNet、SqueezeNet及ResNet50；所有模型中訓練集、驗證集及測試集所占比例分別為63%、27%及10%。使用循序搜尋法(Sequential Search)尋找模型中合適的網路層數、節點個數、Batch大小、Epoch數及學習率，再經混亂矩陣及交叉驗證評估模型；使用準確度最高的模型進行連續多張二維影像之水腦症分類。
ANN分類模型之準確度、靈敏度及特異性分別為88.2%、89.1%及87.9%；ResNet50分類模型之準確度、靈敏度及特異性分別為 96.8%、100%及95.3%；使用ResNet50進行連續兩張影像水腦症之分類準確度為98.6%、靈敏度為100%及特異性為96.8%。
研究結果表明使用人工智慧模型對於水腦症分類是可行的，但解釋性較差，因此可將此人工智慧與定量分析進行互補使用；使用連續兩張二維影像進行判別水腦症不但符合水腦症之生理特性，更可降低單張影像分類錯誤之機率。

關鍵字

水腦症 ； 磁振造影 ； 類神經網路 ； 卷積神經網路

並列摘要

Evan’s Index (EI) and disproportionately enlarged subarachnoid-space hydrocephalus (DESH) are often used as quantitative indicators for the diagnosis of hydrocephalus in clinic. In this study, the investigations of classified hydrocephalus in brain MRI were used artificial intelligent (AI) approaches, including artificial neural network method (ANN) and convolutional neural network (CNN) schemas.
The retrospective study was enrolled in this study. A total of 70 cases were collected, including 31 and 39 cases with respectively to normal and hydrocephalus brain MRI. There were 48 males (51±18.7 yr) and 22 females (49±22.8 yr) with overall average age 50.7 years. The numbers of 2D MRI in normal and hydrocephalus groups were 1073 and 460 images respectively. Three different extracted methods of image features were included in this study, such as five from first-order statistics, twenty from gray-level co-occurrence matrix (GLCM) ,70 from Laws’ texture energy measures (LTEM), and 280 from the combination of GLCM and LTEM. Finally, 40 significant features (P-value < 0.05) were identified by T-test. One ANN model was used to classify normal and hydrocephalus via these 40 features. Meanwhile, three CNN approaches which were AlexNet, SqueezeNet, and ResNet50 were adopted to classify normal and hydrocephalus. The proportions of training and validation and testing sets were 63%, 27% and 10% with respectively to normal and hydrocephalus groups. A sequential searching technique was applied to find the appropriate numbers of layer, number of nodes in each layer, batch size, numbers of epoch and learning rate in these CNN models. These CNN models were evaluated by the confusion matrix and cross-validation in testing set.
The accuracy, sensitivity and specificity were 88.2%, 89.1% and 87.9% generated by ANN with 40 features. The highest accuracy, sensitivity and specificity were 96.8%, 100% and 95.3% generated by ResNet50 with single image among three CNN models. Meanwhile, the accuracy, sensitivity and specificity were 98.6%, 100% and 96.8% provided by ResNet50 with two consecutive 2D images.
ANN and ResNet50 were demonstrated feasibility for classification between normal and hydrocephalus brain MRI. However, the interpretability by ANN and ResNet50 are poor. Hence, the quantitative analytical method might be usefully ancillary the defeat of interpretability for AI models. The ResNet50 was utilized to discriminate Hydrocephalus by using two consecutive 2D images. It not only meets the physiological characteristics of hydrocephalus, but also improves the accuracy of classification.

並列關鍵字

Hydrocephalus ； Magnetic Resonance Imaging ； Artificial Neural Network ； Convolutional Neural Network

參考文獻

[1] Sakka L, Guillaume C, Jean C. Anatomy and physiology of cerebrospinal fluid. European annals of otorhinolaryngology, head and neck diseases. 2011; 128(6):309-316.

[2] Moore KL, Arthur FD, Agur MA. Clinically oriented anatomy. Lippincott Williams & Wilkins. 2013.

[3] Brody H. Medical imaging. Nature. 2013;502(7473):S81-S81.

[4] Kartal MG, Oktay A. Evaluation of hydrocephalus and other cerebrospinal fluid disorders with MRI: an update. Insights into imaging. 2014; 5(4):531-541.

[5] Ng SE, Low AM, Tang KK, Chan YH, Kwok RK. Value of quantitative MRI biomarkers (Evans' index, aqueductal flow rate, and apparent diffusion coefficient) in idiopathic normal pressure hydrocephalus. Journal of Magnetic Resonance Imaging: An Official Journal of the International Society for Magnetic Resonance in Medicine 2009;30(4):708-715.