基於類神經網路之ASPECTS電腦輔助評估研究

Evaluation Study of Neural Networks Based Computer Aids System for ASPECTS

林緯翔(Wei-Hsiang Lin)

指導教授 ： 蘇振隆

中原大學/工學院/生物醫學工程研究所/碩士(2020年)

https://doi.org/10.6840/cycu202000769

摘要

近年來，缺血性腦中風皆高居於我國十大死因第四位，其預後與再灌流治療的實施時間有高度相關。由於腦部斷層掃描的快速性與低成本，臨床診斷中以非對比腦部電腦斷層掃描影像為判讀基礎之ASPECTS評分系統成為一個重要的再灌流治療前評估。而本研究提出一個輔助評估ASPECTS方法，透過可視化缺血性程度協助臨床人員正確且迅速計算ASPECTS分數，於有限時間內有效施打靜脈血栓溶解劑並啟動再灌流治療之評估。
本研究透過影像處理與神經網路技術建構；首先利用雙邊濾波濾除CT影像雜訊，並使用限制對比度的自適應質方圖等化(CLAHE)進行對比度增強。訓練時使用240張腦部CT影像與200張含缺血性病灶影像，透過影像擴增技術放大12000張與16000張對二階段U型全卷積網路耦合隨機條件場模型進行訓練。並以100張DWI-MRI影像病灶位置為標準，採用混淆矩陣相關指標、Dice係數與表面距離指標比較基於系統預測影像進行圈選與醫生在非對比CT影像上圈選病灶區域之相關程度。同時利用圈選結果進行系統評估與醫師評估之ASPECTS分數進行確效。另外透過35組8切面DWI-MRI影像為基準與先前研究進行ASPECTS分數評估，並透過統計10組ASPECTS計算時間來比較新系統與現有系統效能差異。
結果顯示腦組織ROI分割模型其準確率、Dice係數、平均表面距離與最大表面距離達到99.78%、0.994、0.029(pixel)、5.578(pixel)；演算法可正確的分割出腦組織。在缺血性病灶圈選上其Accuracy、敏感度、特異度、交並比、Dice係數、平均表面距離與最大表面距離分別為99.5%、70.1%、99.9、0.82、0.77、1.03(pixel)、15.52(pixel) ；在ASPECTS分數計算上，前腦循環與後腦循環正確率分別達到93%與86%，其圈選的相關度與ASPECTS分數評估正確率皆優於臨床醫師判斷結果；但本系統針對過去既有的缺血區域使用上較有限制。透過神經網絡進行ASPECTS所有切面預測影像可同步輸出，將計算時間縮至30秒左右，在臨床上具有診斷效益。
由結果顯示本研究開發之系統可提供臨床上更快速及準確判斷。透過更多案例之測試，將可實現對急性或超急性腦中風病灶自動化分割及判讀，使系統更具實用性。

關鍵字

急性缺血性腦中風 ； ASPECTS ； 影像處理 ； 全卷積類神經網路

並列摘要

In recent years, ischemic stroke ranks fourth among the top ten causes of death in Taiwan, and its prognosis is highly related to the implementation time of reperfusion therapy. Due to the rapidity and low cost of brain computed tomography (CT), the ASPECTS scoring system based on non-contrast brain CT images in clinical diagnosis has become an important pre-reperfusion evaluation. This study proposes an auxiliary ASPECTS method to assist clinical staff in calculating the ASPECTS score correctly and quickly by visualizing the degree of ischemia, effectively administering venous thrombolytic agents and starting the evaluation of reperfusion therapy within a limited time.

This study uses image processing and neural network technology to construct. First, bilateral filtering is used to filter out CT image noise, and contrast limited adaptive histogram equalization (CLAHE) is used for contrast enhancement. During training, 240 brain CT images and 200 images containing ischemic lesions were used, and 12,000 and 16,000 images were enlarged through data augmentation technology to train the two-stage U-shaped full convolution network (FCN) coupled dense conditional random field(CRF) model. Based on the location of the lesion on 100 DWI-MRI images, the correlation between the confusion matrix related indicators, the Dice coefficient, and the surface distance indicator is compared based on the system prediction image for the circle selection and the doctor's circle selection on the non-contrast CT image. Moreover, the results of ASPECTS scores which calculated from the circle selection compared to the ASPECTS scores evaluated by the physicians are used for system evaluation and validation. In addition, 35 sets of 8 slices DWI-MRI images were used as benchmarks to evaluate the ASPECTS scores with previous studies, and the performance differences between the new and old systems were compared by counting the calculation time of 10 sets of ASPECTS.

The results showed that the accuracy of brain tissue ROI segmentation model, Dice coefficient, average surface distance (ASD), and maximum surface distance (MSD) reached 99.78%, 0.994, 0.029 (pixel), and 5.578 (pixel). The algorithm can segment the brain tissue correctly. The Accuracy, sensitivity, specificity, intersection ratio, Dice coefficient, ASD, and MSD of the ischemic lesion circle are 99.5%, 70.1%, 99.9, 0.82, 0.77, 1.03 (pixel), and 15.52 (pixel), respectively. In the calculation of ASPECTS scores, the accuracy of forebrain circulation and posterior cerebral circulation reached 93% and 86%, respectively, and the correlation between the circle selection and the accuracy of ASPECTS score evaluation were better than the judgment results of clinicians. However, this system is more limited in the use of existing ischemic areas in the past. Through the neural network, all the ASPECTS cross-section prediction images are output synchronously and reducing the calculation time to about 30 seconds, which has clinical diagnostic benefits.

The results show that the system developed by this research can provide faster and more accurate clinical judgment. Through the testing of more cases, the automatic segmentation and interpretation of acute or hyperacute stroke lesions will be realized, making the system more practical.

並列關鍵字

acute ischemic stroke ； ASPECTS ； image processing ； full convolutional neural network

參考文獻

[1] https://www.mohw.gov.tw/cp-16-54482-1.html 108年死因統計結果 (建置日期109-06-16；資料擷取109-07-5)

[2] JS Jeng, TK Lee, YC Chang, ZS Huang, SK Ng, RC Chen, PK Yip, "Subtypes and case-fatality rates of stroke: A hospital-based stroke registry in Taiwan (SCAN-IV)", J Neurol Sci, 156(2):220-6, 1998.

[3] https://dep.mohw.gov.tw/DOS/lp-1777-113.html 104年死因記者會簡報(建置日期105-08-05；資料擷取109-07-16)

[4] JM Wardlaw, J Seymour, J Cairns, S Keir, S Lewis, P Sandercock, "Immediate computed tomography scanning of acute stroke is cost-effective and improves quality of life", Stroke, 35:2477-83, 2004.

[5] HP Adams Jr, TG Brott, RM Crowell, AJ Furlan, CR Gomez, J Grotta, CM Helgason, JR Marler, RF Woolson, JA Zivin, et al.,"Guidelines for the management of patients with acute ischemic stroke. A statement for healthcare professionals from a special writing group of the Stroke Council, American Heart Association", Stroke, 25:1901-4, 1994.