| 摘要: | 背景：腦血管球囊狀動脈瘤是一種病理性的血管突出及擴張，並且可以發生於任何年紀的病人。動脈瘤破裂會造成嚴重的健康問題，例如：出血性中風、腦損傷、昏迷，甚至死亡。腦血管動脈瘤可以在影像檢查中意外被發現，因此，偵測動脈瘤變成放射科醫師一項很重要的任務。然而，近日隨著醫學影像的總量增加，放射線專科醫師的工作量也日漸增加，導致醫師眼力疲乏或專注力降低，造成診斷率下降。因此，此研究欲發展出一套深度學習的模型，預先判斷出腦血管球囊狀動脈瘤的位置，再由放射線專科醫師確認其判讀的正確性，藉此提升動脈瘤的診斷率。
研究目的：發展出一套深度學習的模型，可以自動偵測出核磁共振影像上的腦血管球囊狀動脈瘤，並驗證此模型的診斷率。
材料與方法：此研究回溯性收集臺北市立萬芳醫院從2018-2020年內有進行3D time-of-flight (TOF)磁振血管攝影檢查，並由放射線專科醫師判斷確診為腦血管球囊狀動脈瘤的病例。排除：(1)有蜘蛛膜下腔出血、(2)已接受外科夾除手術（clipping）或血管內介入線圈栓塞（coiling）的患者。從中選出40位患者，將TOF MRA影像去連結後，其中60%病例歸類為訓練群組，20%病例分為驗證群組，另20%則為測試群組，由神經放射線專科醫師利用開源標註軟體（3D Slicer）來進行感興趣影像區域（region of interest, ROI）之正常動脈血管及球囊狀動脈瘤邊界手動圈選，並將這些標註影像匯出NIfTI檔留存。再將已標註之影像，使用DeepMedic模型進行訓練，最後使用測試群組來驗證模型的效能。
結果：本研究在血管分割模型最佳的表現為Dice相似係數0.884，在動脈瘤的偵測上獲得的最佳表現為靈敏度85.7%，陽性預測值37.5%，每例偽陽性1.67顆/位患者。
結論：本研究將深度學習應用於非侵入性的TOF MRA偵測腦部球囊狀動脈瘤，初步的成果可以幫助偵測約80%的動脈瘤，然而，較高的偽陽性率仍是此模型的一大缺點。未來增加訓練資料量及加入正常資料，可以使模型更精準地偵測動脈瘤。
Background: Cerebral saccular aneurysms are outpouching pathological dilatation of intracranial artery. An aneurysm can occur in patients of any age. Ruptured aneurysm can cause serious health problems, such as hemorrhagic stroke, brain damage, coma, and even death. Cerebral aneurysms can be found during image examinations incidentally. Therefore, detection of an aneurysm before it ruptures become an important task for radiologists. However, as the daily image examinations grows recently, radiology reports also increase. The increasing workload of radiologists makes the diagnostic sensitivity decreases due to fatigue of radiologists. Therefore, we want to develop a deep learning model to detect cerebral saccular aneurysm prior to radiologists’ interpretation, to improve the sensitivity of aneurysm diagnosis.
Objectives: This study aims to develop a deep learning model to help radiologists detecting cerebral saccular aneurysms in MRI images and to validate the diagnostic accuracy of this model.
Materials and methods: We retrospectively enrolled and collected patients who had 3D time-of-flight (TOF) MR angiography in Taipei Municipal Wanfang Hospital from 2018 to 2020, and was diagnosed with cerebral saccular aneurysms by radiologist. Exclusion criteria included: (1) patients with subarachnoid hemorrhage, (2) patients who had undergone surgical clipping or endovascular coiling before the image study. We selected 40 patients and de-identify the TOF MRA images. Then, we categorized 60% as the training datasets, 20% as the validation datasets, and 20% as the testing datasets. One neuroradiologist used open-source annotation tool 3D Slicer (https://www.slicer.org) to manually annotate the boundaries of normal arteries and saccular aneurysms, and exported the annotation images as NIfTI files. The annotated images were then used to train the DeepMedic model, and the testing datasets were used to evaluate the model’s performance.
Result: The best performance of the vascular segmentation model in this study was a Dice similarity coefficient of 0.884. The best performance for aneurysm detection achieved sensitivity of 85.7%, positive predictive value of 37.5%, and 1.67 false positives per case.
Conclusion: This study applied deep learning to non-invasive detection of cerebral saccular aneurysms using TOF MRA. The preliminary results show that the model can help detect approximately 80% of aneurysms. However, the relatively high false positive rate remains a significant drawback of this model. Increasing the amount of training data and incorporating normal cases in the future could enable the model to detect aneurysms more accurately. |
