Automatic Detection of Intracranial Hematoma on Routine Brain CT在常規電腦斷層掃描影像中自動辨識顱內血腫

[[abstract]]近年來，不施打顯影劑的非定體積電腦斷層掃描（常規電腦斷層掃描）已成為所有腦神經急症必備的診斷工具，尤其這些疾病多具備突然發生，迅速致命的可怕特性。顱內血腫在斷層掃描影像中呈現高密度（白色）區域，並不難辨認；但不是所有醫師都對複雜腦神經解剖學有相當的了解，要區分各種不同型態的顱內血腫，作出正確的診斷，並不是一件容易的事，而神經內外科或放射線科醫師更無法二十四小時隨侍在側，提供意見，因此病人的診斷與治療可能遭到延誤。 為了克服這個問題，我們參考人類專家的知識，找出必要的規則，發展了從單張常規頭部電腦斷層掃描影像中辨識並區分不同種類顱內血腫的全自動系統，每張影像的辨識可以在一分鐘內完成。我們從一家區域醫院加護病房的連續七十六位病人中，由臨床醫師挑出五十六位病人病灶最明顯的單張斷層掃描影像進行分析。在自發性腦內血腫以及有大型外傷性顱內血腫的影像中，可以得到九成五以上的辨識率；對位在腦內血腫旁邊的腦室內出血以及較小的外傷性顱內血腫，辨識率則較差。 我們發展出的系統，不但具有相當好的敏感度及特異性，影像本身也不須經過任何的前置處理。在結合影像儲存及傳輸系統之後，可望提供臨床醫師即時的決策支援。 In recent years, non-volumetric computed tomography (CT) without administration of intravenous contrast agent has become an important screening tool for almost all patients presenting with acute neurological disorders, which is well known to have great potential of rapid worsening. Intracranial hematoma appears as a hyperdense (white) area on brain CT and is usually not difficult to identify. However, it is not easy to have the correct diagnosis by general physicians due to the complex and usually unfamiliar anatomy. Furthermore, timely opinions from specialist are not usually available, increasing the risk of misdiagnosis and improper treatment. To overcome this problem, we developed a system that can automatically identify and classify different types of intracranial hematoma within single slices of brain CT images in less than one minute, bases on rules derived from human specialists. Fifty-six CT slices bearing the largest hyperdense area of the series were picked up by manually from 76 consecutive patients admitted to the intensive care unit of a single hospital and the results were reviewed. For images with spontaneous deep intracerebral hematoma or large traumatic intracranial hematoma, the system was able to recognize 95% of them (40/42). Worse performance was noted for images containing smaller traumatic hematoma or intraventricular hematoma adjacent to intracerebral hematoma. We conclude that automatic hematoma localization and classification is feasible on single slices of routine non-contrast brain CT for most patients without significant false-positive results. With further refinement, our system may gain wider application and may be integrated into image workstations to form an online decision support system