| 摘要: | 背景：
急性缺血性中風腔隙型梗塞(lacunar infarction)的病人約有20-30%的比例在發病3天內可能發生神經功能缺損惡化，通常稱為早期神經學惡化(early neurological deterioration, END)，可能引起病患失能或是更嚴重的後遺症。血流動力學的灌流不足可能在早期神經學惡化中起關鍵作用，導致梗塞區域生長和功能障礙，甚至長期預後不佳。因此預測哪些病人可能會有灌流不足的風險並儘早開始治療並密切觀察隨時準備再灌流(reperfusion) 非常重要。腔隙型腦梗塞一般認為跟小血管病變的關聯性較大，屬於分支小血管的血流不足引起小範圍的腦梗塞，症狀通常較輕微，但近年來研究指出腔隙型腦梗塞的成因也可能與大血管因素，血栓有關，並不只是單純小血管病變，因此要預測病人是否會有惡化也需考慮更多因素。核磁共振影像廣泛應用於缺血性中風患者，利用擴散加權成像(diffuse weighted image, DWI)提供中風梗塞的體積，利用灌注成像(perfusion-weighted)提供血流動力學狀態及灌流地圖。儘管目前已經提出各種影像以及臨床相關的危險因子與腔隙型腦梗塞的神經學惡化有關，但在識別可以改進預測模型的特徵方面仍然存在挑戰。臨床危險因子各個研究並不一致，一些影像學預測指標缺乏進一步驗證，且更依賴於有經驗的中風專家的判讀，導致臨床應用之困難。依據最近先進的醫學影像研究客觀來說，灌注影像證實腔隙型梗塞如有灌流缺陷，則早期神經學惡化發生的機率較高。我們的目標是擷取核磁共振的影像特徵和機器學習技術來建立灌流缺陷預測模型，以期盡早識別早期神經學惡化的患者。
研究材料與方法:
收案期間自2011年1月至2020年12月回溯收集92位腔隙型梗塞中風病人並在48小時內有進行磁振造影檢查，有臨床檢驗數值，預後追蹤紀錄，有早期神經學惡化紀錄，且有高階灌注成像可證實有無灌流不足。以這些病人最初磁振造影之擴散加權成像(DWI)、擴散係數成像(apparent diffusion coefficient, ADC)、以及液體抑制擴散成像(fluid attenuated inversion recovery, FLAIR)序列擷取影像組學(radiomics)特徵並標準化之後加以篩選，以80%資料作訓練集，10-fold cross validation，保留20%資料作測試集，使用MATLAB軟體利用機器學習建立腔隙型梗塞灌流缺陷預測模型並記錄準確性(accuracy)、特異性(specificity)、敏感度(recall)、(精確率)precision以及F1 值。最後則加上有顯著差異之臨床資訊(腦梗塞體積以及國家衛生研究院中風量表National Institute of Health Stroke Scale, NIHSS score)與影像組學特徵結合建立混合的灌流缺陷預測模型，與前述之模型做比較。
結果:
影像序列分為以下七組進行特徵提取 (DWI、ADC、FLAIR、DWI + ADC、DWI + FLAIR、ADC + FLAIR、DWI + ADC + FLAIR)。各取權重前7之影像組學特徵建立預測模型。其中DWI + FLAIR序列所訓練之模型較佳，準確性為 84.1%，AUROC 0.92，敏感度79.5%，特異性 87.8%，精確率 83.8%, F1 值 81.2。在灌流不足與非灌流不足兩組病患中，有統計顯著差異的臨床因子為NIHSS之分數，及腦梗塞範圍的大小。將此2個特徵與FLAIR序列所得之影像組學前7權重特徵，混合共9項特徵再次利用機器學習建立預測模型，準確性為 88.9%，AUROC 0.91，敏感度 87.5%，特異性 90.0%，精確率 87.5%, F1 score 值。由此可見混合影像組學與臨床特徵的預測模型表現比單純只用影像組學特徵建立之模型要佳。
結論:
使用影像組學技術利用腔隙型梗塞病患腦部核磁共振影像之DWI，ADC，FLAIR序列即可預測其可能存在灌流不足之病況，需密切監測慎防臨床症狀惡化。若混合梗塞體積以及NIHSS則預測模型之表現更佳， 期許可提供第一線人員醫療決策輔助並降低灌流影像成本支出。未來仍需更大型的研究來驗證此成果。
Background:
About 20-30% of patients with acute ischemic stroke, specifically lacunar infarction, may experience worsening of neurological function within 3 days of onset, commonly known as early neurological deterioration (END). This can lead to disability or more severe sequelae in patients. Hemodynamic perfusion insufficiency may play a key role in early neurological deterioration, leading to growth of the infarct area and functional impairment, and even poor long-term prognosis. Therefore, it is very important to predict which patients may be at risk of perfusion insufficiency, to start treatment early, and to closely monitor and prepare for reperfusion. Lacunar infarctions are generally considered to be more associated with small vessel disease, caused by insufficient blood flow in branching small vessels leading to small areas of infarction. Symptoms are usually mild, but recent studies have indicated that the etiology of lacunar infarctions may also be related to large vessel factors and thrombosis, and not merely small vessel disease. Thus, predicting deterioration in patients requires considering more factors. Magnetic resonance imaging (MRI) is widely used in ischemic stroke patients, with diffusion-weighted imaging (DWI) providing the volume of stroke infarction and perfusion-weighted imaging providing the hemodynamic state and perfusion maps. Although various imaging and clinical risk factors related to neurological deterioration in lacunar stroke have been proposed, there are still challenges in identifying features that can improve predictive models. Clinical risk factors vary between studies, some radiological predictive indicators lack further validation, and they rely more on the interpretation of experienced stroke experts, making clinical application difficult. According to recent advanced medical imaging studies, perfusion imaging confirms that lacunar infarctions with perfusion defects have a higher probability of END. Our goal is to extract imaging features from MRI and use machine learning techniques to establish a perfusion defect prediction model, in hopes of identifying patients with END as soon as possible.
Method:
During the period from January 2011 to December 2020, a retrospective collection of 92 patients with lacunar stroke was conducted, who underwent magnetic resonance imaging (MRI) within 48 hours, had clinical laboratory values, follow-up prognosis records, early neurological deterioration records, and advanced perfusion imaging to confirm the presence or absence of perfusion deficits. Using the initial MRI of these patients, radiomics features were extracted and normalized from Diffusion Weighted Imaging (DWI), Apparent Diffusion Coefficient (ADC), and Fluid Attenuated Inversion Recovery (FLAIR) sequences. The data was divided into an 80% training set and a 20% testing set, and a perfusion defect prediction model was developed using machine learning in MATLAB software. Finally, a mixed perfusion defect prediction model was established by combining clinically significant information (stroke volume and National Institute of Health Stroke Scale, NIHSS) with radiomics features, and it was compared with the aforementioned model.
Result:
The feature extraction was conducted on the image sequences, which were divided into seven groups: DWI, ADC, FLAIR, DWI + ADC, DWI + FLAIR, ADC + FLAIR, and DWI + ADC + FLAIR. The top seven radiomic features by weight were selected from each group to construct prediction models. Among these, the model trained on DWI + FLAIR sequence showed superior performance with an accuracy 84.1%，AUROC 0.92，Recall 79.5%，Specificity 87.8%，Precision 83.8%, F1 score 81.2。 Statistically significant clinical factors between patients with and without hypoperfusion included the NIHSS scores and the size of the cerebral infarction. Combining these two features with the top seven weighted radiomics features from the FLAIR sequence, a total of nine features were used to develop a new prediction model through machine learning. This model achieved an accuracy of accuracy 88.9%, AUROC 0.91, Recall 87.5%, Specificity 90.0%, Precision 87.5%, and F1 score 87.5. This indicates that the predictive model integrating both radiomics and clinical features outperforms models constructed with only radiomics features.
Conclusion:
Utilizing radiomics techniques on DWI, ADC, and FLAIR sequences from MRI of patients with lacunar stroke, it is possible to predict the presence of hypoperfusion, necessitating close monitoring to prevent the deterioration of clinical symptoms. Incorporating stroke volume and NIHSS scores into the prediction model enhances its performance, offering a potential tool for frontline medical decision-making and reducing the costs associated with perfusion imaging. Future studies of a larger scale are required to validate these findings. |
