원저접수번호 :09-029(2 차 -0710) 뇌 MRI 를대리결론변수로하는임상시험을위한병변의뇌표준판등록및정량분석소프트웨어개발 : 예비보고 동국대학교일산병원신경과, 인제대학교전산학과 a, 미국라이스대학교생화학세포생물학과 b, 인제대학교의료영상과학과 c 김동억권건환 a 고은아 b 지명구정지원 b 노상미강동희탁윤오 c 김태윤 a 박경종 a 정상욱최흥국 a Development of a Semi-automatic Computer System to Register MRI Lesions Onto a Brain Template for Quantitative Analyses in Clinical Trials Having MRI Findings as Surrogate Endpoints: A Preliminary Report Dong-Eog Kim, MD, PhD, Geon-Hwan Kwan, MS a, Eun Ah Koh b, Myung-Goo Ji, MD, Ji-Won Jeong b, Sang-Mi Noh, MD, Dong-Hee Kang, BS, Yoon-Oh Tak, BS c, Tae-Yun Kim, MS a, Kyoung-Jong Park, BS a, Sang-Wook Jeong, MD, PhD, Heung-Kook Choi, PhD a Department of Neurology, Dongguk University Ilsan Hospital, Goyang, Korea; Department of Computer Science a, Inje University, Gimhae, Korea; Department of Biochemistry and Cell Biology b, William Marsh Rice University, Houston, Texas, USA; Department of Medical Imaging Science c, Inje University, Gimhae, Korea Background: Clinical trials that utilize imaging findings as surrogate endpoints are considered to be cost-effective. However, unlike numeric data, magnetic resonance imaging (MRI) findings are not quantifiable. Thus, we have begun to develop a software package that is able to convert qualitative MRI findings into quantifiable data. Methods: Computer software (DUIH_Image) was created with which every patient s MRI data can be registered on a standard brain template. Interuser and intrauser reliabilities for the registration were measured, and then a proof-of-principle experiment was conducted to determine whether the system could identify factors that were associated with a greater National Institutes of Health Stroke Scale (NIHSS) score at admission. We studied 40 consecutive patients [65.1±14.2 years old (mean±sd); 22 males and 18 females] with first-ever acute lacunar infarction of the corona radiata, who were divided into two groups according to their NIHSS score (i.e., low: 0 2; high: 3). The following parameters were compared between these two groups: (1) data retrieved from clinical profiles, including demographic and risk factor variables; and (2) accumulated diffusion MRI lesions mapped on a standard template. Results: Modest levels of interuser and intrauser reliability were observed (p<0.05, R 2 =0.63 0.84, Pearson correlations). Regarding the clinical profiles, no significant difference was found for the numeric data sets or infarct size between the two groups. However, on the accumulated lesion map image, the lesion area that overlapped the most was located more posterolaterally in the high NIHSS score group than in the low NIHSS score group. Conclusions: In this pilot study we have demonstrated the potential usefulness of the DUIH_Image software. We plan to update this software to enable its utilization in actual clinical trials. J Korean Neurol Assoc 27(4):369-374, 2009 Key Words: Magnetic resonance imaging, Quantification, Computer software, Brain template, NIHSS Received January 13, 2009 Revised June 22, 2009 Accepted June 22, 2009 *Dong-Eog Kim, MD, PhD Department of Neurology, Dongguk University Ilsan Hospital, 814 Siksa-dong, Goyang-si, Gyeonggi-do, 410-773, Korea Tel: +82-31-961-7211 Fax: +82-31-961-7212 E-mail: kdongeog@duih.org 서론 대리결과변수 (surrogate endpoint) 를이용한임상연구는임상결과변수 (clinical endpoint) 를이용한임상연구에비하여시 J Korean Neurol Assoc Volume 27 No. 4, 2009 369
김동억권건환고은아지명구정지원노상미강동희탁윤오김태윤박경종정상욱최흥국 간적, 경제적으로유리하다. 1 그러나, 임상시험에서대규모로얻어지는영상자료는그래픽파일형식 (graphic file format) 으로서그자체로는정성적이기때문에수치자료로이루어진임상정보와달리통계및정량분석에소요되는시간이길어서객관적인분석을신속하게수행하기어려우며, 분석과정에오류가끼어들가능성도상대적으로높다. 뇌표준판은복잡하고크기와모양의개인편차가큰 뇌 들의표준영상으로서고유의 3차원좌표계를가지고있다. 2 따라서, 임상시험피험자의신경영상을뇌표준판으로옮기면정성적인그래픽자료가정량분석이가능한수치자료로변환되는것이다. 3 통계적확률뇌지도 (statistical probabilistic map of brain) 는뇌의특정영역이어떤변수에대하여가지는확률적인정보를제공하는데, 수치의높고낮음에따라서로다른색상-명도- 구배 (color-brightness-gradient) 로표시되도록위색신호 (pseudo-color) 를뇌영상에덧입히는 (overlay) 방법이사용된다. 4,5 임상시험에서군 (group) 간영상자료를비교할때, 뇌표준판에구축된영상자료를토대로확률뇌지도작성기술을이용하여정량분석하게되면수치자료의연산및통계분석에버금가는편의성과정확성을보일것이다. Brain Voyager 등기존의소프트웨어를이용하면뇌MRI 를자동으로뇌표준판에정합 (registration) 시킬수있다. 그러나, 이를위해서는앞맞교차-뒤맞교차 (anterior commissure posterior commissure) 선에평행하게, 그리고 3차원포맷 (three dimensional format) 으로긴시간에거쳐영상을얻어야하기때문에, 신경계환자를대상으로임상에서일반적으로시행되고있는뇌 MRI 에바로적용시키기엔어려운점이많다. 반대로연구자들이뇌병변을수동으로뇌표준판에옮겨그리는방식은시간이많이걸릴뿐아니라그결과물이진정한의미에서정량화된자료라고보기는어렵다. 따라서, 저자들은임상에서일반적으로얻어지는뇌MRI 에서관찰되는병변을특정뇌표준판에반자동적으로정합시킬수있는소프트웨어를개발하고자하였고, 특히신경과전문의가아닌일반연구자나연구보조원도사용할수있도록계획하였다. 본논문에서는상기프로젝트, 즉뇌MRI 소견을대리결론변수로하는임상시험에특화된대규모영상자료구축-자동정량분석장치개발사업의예비결과 (preliminary results) 를보고하고자한다. 대상과방법 뇌MRI 병변을뇌표준판에정합시킴으로써정량화가능한형태로변환저장시켜주는소프트웨어개발을위한개념도는 MRI of the Patient Template MRI Figure 1. Conceptual diagram of MR image registration onto a brain template. Lesions on the MRIs are segmented and registered on the brain template (arrows). Fig. 1과같다. 이소프트웨어 (DUIH_Image) 는 Visual C++6.0 을이용하여개발되었으며, 공개라이브러리인 CxImage 6,7 와 GDI+ 8 를사용하였다. 개발된소프트웨어인 DUIH_Image 의실행가능성을평가하기위한첫단계로서원리증명실험 (proof-of-principle experiment) 을수행하였다. 2007 년 1월부터 2008 년 8월까지동국대학교일산병원에입원한 1주일이내급성뇌경색환자 419 명중뇌졸중이재발한경우가아닌환자로서, 확산강조영상에서대뇌부챗살 (corona radiata) 에열공성뇌경색이관찰되고, TOAST (Trial of Org 10172 in Acute Stroke Treatment) 9 분류상소혈관질환을기전으로하는 40명 (65.1±14.2 세, 남자 22명, 여자 18명 ) 을대상으로하였다. 두명의연구보조원이 DUIH_Image 를이용하여확산 -강조 MRI 에서관찰되는급성뇌경색병변을뇌표준판에정합시키는작업을수행하였고, 신경과전문의의검토를받았다. 병변을뇌표준판에정합시킬때우뇌의병변은좌뇌로대칭이동시켰다. 본예비연구에서는뇌표준판으로서이번연구에포함되지않은한환자의액체감약반전회복 (fluid attenuation inversion recovery, FLAIR) MRI 를사용하였다. 사용자간 (inter-user; 연구보조원과신경과전공의 ) 및사용자내 (intra-user; 신경과전공의 ) 정합일치도를측정하기위하여, 무작위로선정된 9명의환자로부터다시무작위로선정된 18개병변의무게중심 (center of gravity) 이선형상관관계를보이는지검증하였다. 사용자내정합일치도판정을위해서는같 370 대한신경과학회지제 27 권제 4 호, 2009
뇌 MRI 를대리결론변수로하는임상시험을위한병변의뇌표준판등록및정량분석소프트웨어개발 : 예비보고 은영상자료에대해 3개월간격을두고반복하여정합작업을수행하였다. 신경해부학적지식과기존의연구결과 10,11 를바탕으로다음과같은선험적전제 (a priori premises) 및가설하에영상분석을시작하였다. 즉, 연구대상환자를입원 NIHSS (National Institute of Health Stroke Scale) 점수의중앙값을기준으로신경학적손상도가낮은군 (group) 과높은군으로양분하였을때, 그룹별로해당환자의뇌경색병변을누적시켜지도화해보면가장많이겹치는열점 (hot spot) 의위치가두군간에차이를보일것이다. 이와더불어병변의크기 ( 픽셀수 ), 대뇌부챗살뇌경색이섬유막및기저핵영역까지침범하고있는증례의빈도, 전향적으로조사된인구학인자 (demographic variables) 및혈관계위험인자등이두군간에차이를보이는지도분석하였다. 내원전에병의원에서고혈압으로진단받은경우또는고혈압약약을투여받은병력, 즉과거력이있는경우고혈압이있는것으로판정하였다. 입원이후뇌졸중회복기 (1~2 주이상시간경과 ) 에혈압이 140/90 mmhg 이상으로높아서고혈압약을복용해야하는경우또는고혈압에의한장기손상 (end-organ damage) 의증거가있는경우에도고혈압으로판정하였다. 당뇨는과거력이있거나, 뇌졸중회복기에도공복혈당이 126 mg/dl 이상이거나식후 2시간째혈당이 200 mg/dl 이상인경우로정의하였다. 고지혈증은과거력이있거나, 입원직후 8시간이상공복후시행한검사중 1회이상 LDL (low density lipoprotein) 콜레스테롤치가 160 mg/dl 이상으로나왔을때또는총콜레스테롤수치가 240 mg/dl 이상일때로정의하였다. 입원혈압은내원하여측정한첫번째혈압을기록하였다. 통계학그룹간평균의비교는 Student s t 검정을사용하였고, 분율의비교는카이제곱검정을사용하였다. 그리고, 선형상관관계를확인하기위하여피어슨상관계수를사용하였다. p 값이 0.05 미만인경우통계적으로유의하다고판정하였다. 결과 개발된 DUIH_Image 는다음과같이작동하였다 (Fig. 2). 읽어들인뇌MRI 에서관심영역을분할하기위한첫단계로분할뷰어 (segmentation viewer) 창을통해두개골또는뇌의외곽선을자동추출시키는작업을거치도록되어있는데, 이과정에서 Otsu 연산법 12 이사용된다. 분할크기 (segmentation Size) 항 Figure 2. Operation of DUIH Image. If a user clicks the Auto Outline button, external surface of the skull is automatically outlined in red color (A). By adjusting the Black-White scale bar and selecting the Segmentation Size (A), high or low intensitity pixels are labeled with yellow color (B). By clicking the yellow areas, regions of interest are selected in green color (C). After automatic registration onto the template (yellow color), semi-atuomatic adjustment of the location and size is available, allowing a user to refine the registration (D). Once registered, the MR data become quantifiable for various statistical analyses (E). 목과흑백눈금막대 (Black-White scale bar) 를이용하면밝기 (brightness) 가상대적으로높거나낮은부분을자동으로레이블링할수있다. 자동레이블링된영역중관심영역을마우스로클릭하면녹색으로표시된다. 선택된관심영역을특정뇌표준판에자동으로등록시킴으로써정량화가능한자료로변환시키는정합과정을거치게된다. 자동으로정합된관심영역에 J Korean Neurol Assoc Volume 27 No. 4, 2009 371
김동억권건환고은아지명구정지원노상미강동희탁윤오김태윤박경종정상욱최흥국 Figure 3. Inter-user correlation and intra-user correlation. In the upper row, there is a significant linear correlation (p<0.05, Pearson correlation) of x / y coordinates of the centers of the stroke lesions (n=18 from randomly selected 9 patients) between the research assistants registration (horizontal-axis) and a neurology resident s registration (vertical-axis). In the lower row, there is a significant intra-user (neurology resident) correlation (p<0.05, Pearson correlation). 대해원본과비교하여왜곡되었다고판단되는경우사용자가위치변경, 확대, 축소등미세조정을수행할수있도록하는반자동교정장치가내재되어있다. 이러한과정을거쳐뇌표준판으로정합된자료들이저장되면, 언제든지연구하고자하는기준에맞는환자들의자료를불러들여병변을중첩시켜가며정량분석할수있다. 사용자간및사용자내정합일치도를검증한결과통계적으로유의한선형상관관계를확인할수있었다 (p<0.05, 피어슨계수, R=0.79~0.92). 결정계수 (R 2 ) 는 0.63~0.84 로계산되었다 (Fig. 3). 원리증명실험에서 NIHSS 점수의중앙값을기준으로신경학적손상도가낮은 (low NIHSS) 군과높은 (high NIHSS) 군으로양분하였을때, 문자 / 수치자료인인구통계적인자나혈관계위험인자에있어서는두그룹간에통계적으로유의한차이를보이지않았다 (Table). 뇌경색의크기나대뇌부챗살뇌경색이섬유막및기저핵영역까지침범하고있는증례의빈도에있어서도차이를보이지않았다 (Table). 그러나, DUIH_Image 를이용하여연구대상환자들의뇌경색병변을누적시켜지도화했을때, 4 슬라이스중 3 슬라이스에서열점의위치가 NIHSS 가높은군에서낮은군에비해좀더후외측에위치함을알수 있었다 (Fig. 4). 병변의무게중심 (center of gravity) 만을 x-y 좌표상에표시하였을때, DUIH_Image 를이용해병변전체를누적지도화시킨결과와부합됨을확인할수있었다 (Fig. 4). 고찰 2007 년 6월기준, 전세계임상시험시장규모는약 40조원에육박한다. 국가별로미국, 캐나다, 호주등에이어한국은임상시험점유율이 1.4% 로 9위에위치하고있으며이를금액으로환산시키면약 1,000 억원규모로추계된다. 주목할것은신약개발에있어제1~3 상에이르는단계별시간과비용이점차급증하고있다는점이며, 2001 년을기준으로보면 5년전에비하여 50% 증가하였다. 임상적결론변수를사용하는일반적인임상시험에비해영상대리결론변수를사용하는임상시험은상대적으로비용이덜드는경우가많다. 1 신경영상을이용한척도는임상척도에비해대조군과치료군의차이를더민감하게평가할수있어연구대상자수와관찰기간을줄일수있기때문이다. 뇌MRI 소견을대리결론변수로이용하는임상시험의또다른장점은투여하고자하는약물의기전에적합한대상환자를선택할수있고, 약물기전의효과를다양한영상기법을이용 372 대한신경과학회지제 27 권제 4 호, 2009
뇌 MRI 를대리결론변수로하는임상시험을위한병변의뇌표준판등록및정량분석소프트웨어개발 : 예비보고 Table. Demographic and Clinical Variables Low (0~2) N = 17 Admission NIHSS High ( 3) N = 23 Median Admission NIHSS 1 3 Age 64.2±13.4 65.1±14.5 0.83 Sex, Male 10 (58.9%) 14 (60.9%) 0.90 Hypertension 12 (70.6%) 14 (60.9%) 0.52 Diabetes mellitus 3 (17.6%) 4 (17.4%) 0.98 Hyperlipidemia 4 (23.5%) 7 (30.4%) 0.63 Smoking 9 (52.9%) 11 (47.8%) 0.75 Total cholesterol (mg/dl) 175.2±30.0 175.6±30.0 0.96 High density lipoprotein cholesterol (mg/dl) 41.3±12.2 43.8±10.3 0.53 Low density lipoprotein cholesterol (mg/dl) 100.8±21.3 105.5±30.9 0.62 Triglyceride (mg/dl) 140.8±71.4 132.6±75.2 0.75 Admission systolic blood pressure (mmhg) 161.9±29.1 166.1±21.9 0.60 Admission diastolic blood pressure (mmhg) 95.5±16.0 92.2±12.7 0.48 Lesion extension to the capsular area 9 (52.9%) 13 (56.5%) 0.82 Lesion size (pixel counts) 472.0±298.9 501.8±458.5 0.83 NIHSS denotes NIH stroke scale. The Student s t-test was used for comparison of continuous variables between groups. The Chi-Square test was used for comparison of proportions. p value 해검증할수있다는점이다. 1 미식약청의 1997 년결의서 (FDA Modernization Act of 1997) 에의하면신경보호약물의경우한개의 3상임상시험 결과와한개의 MR 뇌영상소견을대리결론변수로하는 2상임상시험 결과만있으면신약허가를신청할수있다. 13 뇌MRI 소견을대리결론변수로이용하기위해서는분석의정확도가높아야한다. 이를위해서는임상시험에서얻어지는대규모영상자료로부터특정자료를추출하고이영상자료를수치자료와다름없이쉽게정량 / 통계분석할수있는도구가필요하다. 그런데, 임상시험에 특화 된영상자료구축 -자동분석장치는아직없는상태이다. 이러한장치의개발은임상시험의분석정확도를높일뿐아니라경제성과편의성을함께향상시킴으로써뇌영상소견을대리결론변수로하는연구가더많이수행되도록하는촉매역할을할수도있을것이다. 그첫단계로서저자들은임상에서일반적으로얻어지는뇌 MRI 에서관찰되는병변을뇌표준판에반자동적으로정합시킬수있는소프트웨어의베타버전을개발하였다. 그리고, 개발된소프트웨어를사용하여신경과전문의가아닌연구보조원이연구대상환자 40명의확산강조영상급성열공성대뇌부챗살뇌경색 (corona radiata infarction) 병변을뇌표준판에손쉽게등록, 정합시킬수있었다. 정합된자료를신경과전문의가검토한결과병변의크기와위치에대한수정이필요한경우는없 었다. 그리고, 이자료를이용하여뇌경색병변을누적시켜지도화했을때열점의위치, 즉병변이가장빈번하게관찰된영역이 NIHSS 점수가높은군에서낮은군에비해좀더후외측에위치함을알수있었는데, 이것은기존의연구결과 10,11 와잘부합하는소견이다. 대뇌부챗살과섬유막및기저핵영역을지나는운동섬유로 (pyramidal motor tract) 와시상-피질감각섬유로 (thalamo-cortical sensory tract) 가주로뇌의중앙수평선근처에서약간뒤쪽으로치우쳐위치하기때문에, 14 상대적으로뒤쪽에위치한열공성뇌경색병변을가지는환자들에서 NIHSS 점수가더높았을것이다. 한편, 정량화된영상자료와달리문자 / 수치자료인인구통계적인자나혈관계위험인자모두에있어서는 NIHSS 점수가높은군과낮은군사이에통계적으로유의한차이가관찰되지않았다. 이결과는뇌경색환자의신경학적결손도를결정하는요인으로서병변의위치를강조한 Menezes 등의연구 15 와맥락을같이하는것이다. 또한, 본연구를통해제작된뇌 MRI 정량화소프트웨어인 DUIH_Image 가임상연구에서유용한도구로사용될수있음을시사하는것이다. 그러나, DUIH_Image 가실제임상시험또는뇌MRI 관련임상연구에서널리사용되기위해서는다음과같은과정을거쳐야한다. 첫째, 베타버전에비해안정성과사용자편의성을높인정식버전으로의향상 (upgrade) 이필요하다. 둘째, 사용자간 J Korean Neurol Assoc Volume 27 No. 4, 2009 373
김동억권건환고은아지명구정지원노상미강동희탁윤오김태윤박경종정상욱최흥국 ACKNOWLEDGEMENTS 본연구는보건복지부보건의료기술진흥사업의지원에의하여이루어진것임 ( 임상시험기술개발사업 ; A070001). REFERENCES Figure 4. Probabilistic accumulation lesion map and scatter plot for the centers of the stroke lesions in the low (0~2) and high ( 3) NIHSS groups. In the representative accumulated diffusion MR lesion maps on a template, the lesion area overlapped the most is located more posterolaterally in the high NIHSS group than in the low NIHSS group. This finding is corroborated by the scatter plot for the coordinates of the centers of each individual stroke-lesion in the low and high NIHSS groups. The scales on the x and y axes cover the left corona radiata and capsular area. Small cross-hairs in the lesion maps correspond to the big cross-hair in the scatter plot. 및사용자내정합일치도가상관관계결정계수기준으로 0.63~ 0.84 정도여서개선된방식의정합-알고리즘이탑재되어야할것으로판단된다. 셋째, 뇌MRI 병변을뇌표준판으로정합시키는데필요한시간과노력이일반적인연구팀에서받아들일수있는수준이어야하겠다. 저자들은뇌MRI 자료를정량적으로구축하고분석하는데사용할수있는컴퓨터소프트웨어를개발하였으며, 원리증명실험을통해유용성을확인한초기결과를얻었기에이를보고하는바이다. 1. Hong KS. Outcome Measure and Efficacy Analysis in Stroke Clinical Trials. J Korean Neurol Assoc 2006;24:411-420. 2. Toga AW, Thompson PM. Maps of the brain. Anat Rec 2001;265: 37-53. 3. Lee JS, Lee DS, Kim YK, Kim JS, Lee JM, Koo BB, et al. Quantification of Brain Images Using Korean Standard Templates and Structural and Cytoarchitectonic Probabilistic Maps. Korean J Nucl Med 2004;38:241-252. 4. Evans A, Collins D, Holmes C. Automatic 3D regional MRI segmentation and statistical probability anatomy maps. In: Myers R. Cunningham V. Bailey D, Jones T, eds. Quantification of brain function: tracer kinetics and image analysis in brain PET. London: Academic Press, 1996;123-130. 5. Koo BB, Lee JM, Kim JS, Lee JS, Kim IY, Kim JJ, et al. Developing a Korean Standard Brain Atlas on the basis of Statistical and Probabilistic Approach and Visualization tool for Functional image analysis. Korean J Nucl Med 2003;37:162-170. 6. D. Pizzolato. CxImage. 2003-2008 [cited 2009 April 1; Available from: URL: http://www.xdp.it/cximage.htm 7. Jeong SH, Lee MH. Visual C++ Digital imaging using the open source CxImage. Seoul: Hongrung publishing company, 2006;457-577. 8. Chand M. Graphics programming with GDI+. Reading: Addison- Wesket, 2004;273-523. 9. Adams HP Jr, Woolson RF, Clarke WR, Davis PH, Bendixen BH, Love BB, et al. Design of the Trial of Org 10172 in Acute Stroke Treatment (TOAST). Control Clin Trials 1997;18:358-377. 10. Kim JS, Pope A. Somatotopically located motor fibers in corona radiata: evidence from subcortical small infarcts. Neurology 2005;64: 1438-1440. 11. Song YM. Somatotopic organization of motor fibers in the corona radiata in monoparetic patients with small subcortical infarct. Stroke 2007;38:2353-2355. 12. Parker JR. Algorithms for image processing and computer vision. New York: John Wiley & Sons Inc, 1997:116-149. 13. Hutt PB. A guide to the FDA Modernization Act of 1997. Food Technology 1998;52:54. 14. Yamada K, Nagakane Y, Yoshikawa K, Kizu O, Ito H, Kubota T, et al. Somatotopic organization of thalamocortical projection fibers as assessed with MR tractography. Radiology 2007;242:840-845. 15. Menezes NM, Ay H, Wang Zhu M, Lopez CJ, Singhal AB, Karonen JO, et al. The real estate factor: quantifying the impact of infarct location on stroke severity. Stroke 2007;38:194-197. 374 대한신경과학회지제 27 권제 4 호, 2009