Analysis of head CT scans flagged by deep learning software for acute intracranial hemorrhage

Neuroradiology - To analyze the implementation of deep learning software for the detection and worklist prioritization of acute intracranial hemorrhage on non-contrast head CT (NCCT) in various...     

Preliminary interpretations of after-hours CT and sonography by radiology residents versus final interpretations by body imaging radiologists at a level 1 trauma center

OBJECTIVE: At many academic institutions, preliminary interpretations of CT scans and sonograms obtained after regular hours of operation are performed by radiology residents, with attending radiologists reviewing the interpretations the next morning. We sought to determine the rate of discrepancy between residents' interpretations of imaging studies and the final interpretations performed by an attending body imaging radiologist as well as any resulting clinical consequences stemming from the discrepancies. Therefore, we reviewed 928 CT and sonographic images that had been obtained after hours at a level 1 trauma center during a 6-month period. MATERIALS AND METHODS: Any discrepancies between the preliminary and final interpretations were judged as either major (i.e., necessitating an urgent change in treatment) or minor errors. We conducted patient follow-up via a retrospective review of the medical charts to determine whether any of the discrepancies led to additional imaging, an increase in patient morbidity, an extension of a hospital stay, or a change in treatment. RESULTS: The overall discrepancy rate in interpretations rendered by the residents and those performed by the attending radiologist was 3.8%, with most of these discrepancies (86%) judged to be minor. If we combined the data for body CT scans and sonograms, the rate of minor discrepancies was 3.2%, and the rate of major discrepancies was 0.5%. If we considered only body CT data in the evaluation, the overall discrepancy rate increased to 6.4%, with a 5.4% rate of minor discrepancies and a 1.0% rate of major discrepancies. CONCLUSION: Our evaluation of discrepancy rates was unusual in that we included interpretations of sonograms, on which residents and the attending radiologist had a higher rate of agreement (99.5%). Because of the high agreement in the interpretation of sonograms, the overall discrepancy rate was 3.8%. However, if only body CT scan interpretations were evaluated, our results were closer to the rates reported in previously published studies. Major discrepancies led to a change in patient treatment but did not lead to any increase in patient morbidity or to any quantifiable increase in the length of the hospital stay.     

CT detection of intracranial aneurysms in subarachnoid hemorrhage.

Contrast enhanced computed tomography (CT) scans of 59 patients with subarachnoid hemorrhage were evaluated for intracranial aneurysms. After prospective study, the CT results were compared with angiography. Depending on the location, the CT detection rate varied from a low of 36% for the internal carotid artery complex to a high of 76% for anterior and middle cerebral artery aneurysms. It was concluded that angiography may be deferred if the causative aneurysm is demonstrated by CT.     

Acute intracranial hemorrhage secondary to thrombocytopenia: CT appearances unaffected by absence of clot retraction.

PURPOSETo describe the in vivo CT appearance of acute intracerebral blood clots formed from anemic platelet-depleted blood.METHODSThree patients with intracerebral hemorrhage secondary only to thrombocytopenia were examined with CT within 2 1/2 hours after the onset of clinical symptoms.RESULTSThere were no unusual CT features found in the intracerebral hemorrhages of patients with only thrombocytopenia. Specifically, a hyperdense zone(s) surrounded by areas of decreased density was identified.CONCLUSIONClot retraction (which cannot occur in patients with severe thrombocytopenia) is not necessary for the CT appearance of acute intracerebral hemorrhage.     

双源CT虚拟平扫在头部检查的初步应用

目的 探讨双源CT双能量头部虚拟平扫(VNC)的图像质量和临床应用价值.方法 对62例临床怀疑脑血管病变的患者,使用双源CT进行头部常规平扫(CNC)和双能量CTA检查,利用双能量软件得到VNC图像.比较CNC和VNC图像灰质、白质、脑脊液、高密度出血性和低密度缺血性病变的CT值,使用4分法对图像质量进行主观评价,比较两组图像的噪声、辐射剂量和病变检出率,使用配对t检验、Wilcoxon符号秩检验和χ2检验(McNemar检验和Kappa检验)进行统计分析.结果 CNC与VNC图像灰质、白质、脑脊液、高密度病变及低密度病变的CT值分别为[(43.3±1.5)和(33.2±1.3)HU,t=46.98]、[(32.9±1.3)和(28.8±1.6)HU,t=16.28]、[(9.0±1.4)和(5.3±1.9)HU,t=12.41]、[(62.8±10.0)和(51.3±11.5)HU,Z=-4.37]、[(20.7±4.7)和(18.0±6.9)HU,t=3.84],差异均有统计学意义(P值均＜0.01).VNC图像噪声[(1.63±0.34)HU]大于CNC图像[(0.99±0.18)HU](Z=-6.41,P＜0.01).VNC图像有效剂量[(0.53±0.08)mSv]低于CNC[(1.37±0.23)mSv](Z=-6.45,P＜0.01).CNC和VNC图像噪声、颅底伪影、灰白质对比、高密度和低密度病变显示的主观评分分别为(3.9±0.3)和(2.7±0.5)分、(3.7±0.5)和(2.4±0.9)分、(3.3±0.6)和(1.3±0.5)分、(4.0±0.0)和(3.0±0.4)分、(3.9±0.3)和(3.2±0.8)分,VNC图像噪声与颅底伪影的评分较CNC图像低(Z值分别为-6.84、-6.15,P值均＜0.01),灰白质对比、高密度和低密度病变显示低于CNC图像(Z值分别为-7.01、-4.52和-3.12,P值均＜0.01).在个体水平,VNC图像显示高密度出血性病变29例,无假阳性和假阴性病例,敏感性、特异性、阳性预测值和阴性预测值均为100.0%(29/29、33/33、29/29、33/33),与CNC一致(P＞0.05,Kappa值=1.000);VNC图像显示低密度缺血性病变22例,假阳性1例,假阴性2例,敏感性、特异性、阳性预测值和阴性预测值分别为91.3%(21/23)、97.4%(38/39)、95.5%(21/22)和95.0%(38/40),与CNC图像(23例)差异无统计学意义(χ2=0.00,P＞0.05,Kappa值=0.895).在病灶水平,VNC图像显示出血灶53个,假阴性4个,无假阳性,敏感性、特异性、阳性预测值和阴性预测值分别为93.0%(53/57)、100.0%(38/38)、100.0%(53/53)和90.5%(38/42),VNC图像对出血灶的显示率与CNC差异无统计学意义(χ2=2.25,P＞0.05,Kappa值=0.914);VNC图像显示低密度病灶38个,假阳性2个,假阴性13个,敏感性、特异性、阳性预测值和阴性预测值分别为73.5%(36/49)、96.4%(53/55)、94.7%(36/38)和80.3%(53/66),VNC图像对低密度病灶的显示率低于CNC(χ2=6.67,P＜0.01,Kappa值=0.707).结论 与CNC相比,头部VNC辐射剂量低,但图像质量下降,对出血性病变具有替代CNC的潜在使用价值,对缺血性病变也有一定的参考价值.

Abstract：

Objective To investigate image quality and clinical value of dual-source dual energy virtual non-contrast (VNC) CT of the head. Methods Sixty-two patients suspected of cerebrovascular diseases underwent conventional non-contrast (CNC) CT and dual energy CTA examination of the head with dual-source CT. Virtual non-contrast images were reconstructed using dual energy software. The CT values of gray matter, white matter, cerebrospinal fluid, hyperdense hemorrhagic lesion and hypodense ischemic lesion were compared between CNC and VNC images. A four-score scale was used to assess image quality subjectively. Image noise, radiation dosage and detection rate were compared between CNC and VNC images. Paired t test, Wilcoxon signed ranks test and Chi-square test (McNemar test and Kappa test) were used. Results The CT value on CNC and VNC images, were (43. 3 ± 1.5) and (33. 2 ± 1.3) HU for gray matter (t = 46.98, P ＜ 0. 01), (32. 9 ± 1.3) and (28.8 ± 1.6) HU for white matter(t = 16. 28, P ＜0.01), (9.0 ± 1.4) and (5.3 ± 1.9) HU for cerebrospinal fluid (t=12.41, P＜0.01),(62.8 ±10.0) and (51.3 ± 11.5) HU for hyperdense lesion (Z = -4.37, P ＜ 0.01), (20.7 ±4.7) and (18.0 ±6. 9) HU for hypodense lesion (t = 3. 84, P ＜ 0. 01), respectively. VNC images[(1.63 ±0.34) HU]had more noise than CNC images[(0.99±0.18) HU](Z= -6.41, P＜0.01). VNC [(0. 53 ± 0. 08) mSv]had less effective dose than CNC[(1.37 ± 0. 23) mSy](Z= - 6. 45, P ＜ 0. 01).In subjective assessment, VNC images had more noise (2. 7 ± 0. 5 for VNC and 3.9 ± 0. 3 for CNC,Z = -6. 84, P ＜ 0. 01) and skull base-related artifacts (2. 4 ± 0. 9 for VNC and 3.7 ± 0. 5 for CNC,Z = -6. 15, P ＜0. 01) than CNC images. The gray/white matter contrast (1.3 ± 0. 5 for VNC and 3.3 ±0. 6 for CNC, Z = - 7. 01, P ＜ 0. 01), hyperdense lesion display (3.0 ± 0. 4 for VNC and 4. 0 ± 0. 0 for CNC,Z = -4. 52, P ＜ 0. 01) and hypodense lesion display (3.2 ± 0. 8 for VNC and 3.9 ± 0. 3 for CNC,Z= -3. 12, P ＜0. 01) on VNC images were lower than those on CNC images. In per-patient analysis,29 cases of hyperdense lesion (hemorrhage) were found on VNC images without misdiagnosis. The sensitivity, specificity, positive predictive value and negative predictive value were all 100. 0% (29/29,33/33, 29/29, 33/33). VNC images had the same detection rate of hyperdense lesions as CNC images (P ＞0. 05, Kappa = 1. 000) at per-patient level. Twenty-two patients with hypodense ischemic lesions were found on VNC images with one false positive case and two false negative cases. The sensitivity,specificity, positive predictive value and negative predictive value were 91.3% (21/23), 97.4%(38/39), 95.5% (21/22) and 95.0% (38/40) respectively. No statistical difference was found in detecting hypodense lesions between VNC and CNC images (χ2 = 0. 00, P ＞ 0. 05, Kappa = 0. 895). In per-lesion analysis, 53 hemorrhage lesions were found on VNC images with false negative results of four lesions and no false positive result. The sensitivity, specificity, positive predictive value and negative predictive value were 93.0% (53/57), 100. 0% (38/38), 100. 0% (53/53) and 90. 5% (38/42)respectively. There was no significant difference in detection rate of hyperdense lesion between VNC and CNC images (χ2 =2. 25, P ＞0. 05, Kappa =0. 914). Thirty-eight hypodense lesions were found on VNC images with 2 false positive lesions and 13 false negative lesions. The sensitivity, specificity, positive predictive value and negative predictive value were 73.5% (36/49), 96.4% (53/55), 94. 7% (36/38)and 80. 3% (53/66) respectively. The detection rate of hypodense lesion on VNC images was lower than that on CNC images (χ2 = 6. 67 ,P ＜ 0.01, Kappa = 0. 707). Conclusion Compared with CNC images,head VNC images have reduced image quality and radiation dosage. VNC images can replace CNC images potentially in detecting intracranial hemorrhage and provide information for ischemic cerebrovascular diseases to some extent.     

颅脑外伤性迟发性出血的CT诊断

目的：用ＣＴ随访研究颅内迟发性血肿,方法：在２３２例颅脑外伤患者中行ＣＴ随访检查,结果：６１例发现有迟发性颅内出血,其中３８例为原血肿增大或出现新的血肿;８例为对冲伤血肿,１５例为脑挫伤与小量散在出血灶,结论：在外伤病例中,即使开初ＣＴ扫描为阴性者,如病人的临床表现有改变,如症状加重,意识丧失,抽搐等,应立即行ＣＴ复查,对ＣＴ随访的价值及迟发出血的机理也进行了讨论。     

颅内肿瘤出血CT诊断(附57例分析)

颅内肿瘤出血是颅内肿瘤急性发病期主要的致死原因之一,文献报道颅内肿瘤出血的发生率在3%～10%之间.本文收集1990年1月～2000年2月57例资料完整的颅内肿瘤出血病人,均经CT及手术病理证实.     

CT detection and course of intracranial hemorrhage in premature infants.

Twenty neonates with a suspected intracranial hemorrhage were studied by computed tomography (CT). The exact site and extent of the hemorrhage in all infants were clearly demonstrated on serial CT scans. In intraventricular hemorrhage, a dense subependymal halo lined the ventricular system and could be recognized for up to 2 weeks. Discrete hemorrhage adjacent to the ventricular system also appeared as discrete nodules rather than as a diffuse hemorrhage. Blood in the ventricular system could be recognized up to 2 weeks when there were blood-cerebrospinal fluid levels. Hydrocephalus was a common sequela and was readily detectable before a measurable change in head size.     

Spiral head CT in the evaluation of acute intracranial pathology: a pictorial essay

The purpose of this pictorial essay is to review the benefits of spiral head computed tomography (CT) with routine multiplanar reformations in the evaluation of acute intracranial pathology. This technique is particularly useful in trauma patients for detection of skull base or calvarial fractures, thin tentorial subdural hematomas, or for more specific characterization of intracranial hemorrhage. The benefits of multiplanar reformations have been described for a variety of other diagnoses in the chest, abdomen, extremities, and spine, and their routine use continues to grow with the widespread availability of multi-slice CT scanners. In this article, we describe spiral head CT technique with multiplanar reformations as an alternative to the routinely used sequential technique. Subtle findings and lesions aligned predominantly in the axial plane can often be visualized to better advantage with multiplanar reformations. We also address technical factors for optimizing spiral technique.     

CT patterns of intracranial hemorrhage complicating thrombolytic therapy for acute myocardial infarction

Computed tomographic (CT) patterns of intracranial hemorrhage (ICH) were determined in 1,696 patients undergoing thrombolytic therapy for acute myocardial infarction. ICH occurred at 33 sites in 0.77% of patients (n = 13). Thirty-six percent of hemorrhages (n = 12) were intraparenchymal, 33% (n = 11) were subdural, 24% (n = 8) were subarachnoid, and 6% (n = 2) were intraventricular. Eight-four percent (n = 26) of all nonventricular hemorrhages were supratentorial in location. The most common site of ICH was supratentorial and intraparenchymal (10 of 33). In 11 of the 13 patients with ICH, clinical symptoms occurred within 24 hours of the initiation of thrombolytic therapy. A fatal outcome resulted in three of the four patients who had clinical symptoms within 3.5 hours after beginning thrombolytic therapy. The CT findings of multiple intracranial bleeding sites, substantial mass effect with midline shift, and large-volume intraparenchymal hematomas were associated with increased mortality. ICH associated with thrombolytic therapy for acute myocardial infarction has a grave prognosis, with 62% of patients dying during hospitalization.     

Flat-panel detector volumetric CT for visualization of subarachnoid hemorrhage and ventricles: preliminary results compared to conventional CT

INTRODUCTION: The aim of this study was to compare flat-panel volumetric CT (VCT) to conventional CT (cCT) in the visualization of the extent of subarachnoid hemorrhage (SAH) and the width of the ventricles in patients with acute SAH. METHODS: Included in the study were 22 patients with an acutely ruptured cerebral aneurysm who received VCT during coil embolization. VCT image quality, the extent of SAH (using a modified Fisher score and total slice number with SAH visible) and the width of the ventricles (Evans index) were evaluated by two experienced neuroradiologists (RAD1 and RAD2) and compared to the findings on cCT. Ten patients undergoing VCT for reasons other than SAH served as negative controls. RESULTS: Interobserver agreement in rating image quality was excellent for cCT (Kendall W value 0.94) and good for VCT (0.74). SAH was identified by RAD1 and RAD2 on VCT images in all patients. The modified Fisher scores underestimated the extent of SAH on VCT images in comparison with cCT images. Pearson's correlation coefficient (r) regarding the number of image slices with SAH visible on cCT images compared with the number on VCT images was 0.85 for RAD1 and 0.84 for RAD2. The r value for the degree of interobserver agreement for the number of slices with SAH visible was 0.99 for cCT, and 0.95 for VCT images (n=19), respectively. The width of the ventricles measured in terms of the Evans Index showed excellent concordance between the modalities (r=0.81 vs. 0.82). CONCLUSION: Our preliminary results indicate that VCT is helpful in evaluating SAH in the angiography suite. Additionally, reliable evaluation of ventricle width is feasible. However, there are limitations with regard to the visibility of SAH on VCT images in comparison to cCT images.     

颅内介入术后即刻头颅CT高密度影像分析

目的：分析颅内介入术后（包括动脉瘤栓塞、颅内支架植入及急诊动脉溶栓术）即刻CT颅内高密度灶的类型及其对临床治疗的指导意义。方法：回顾性分析103例患者（其中动脉瘤栓塞术37例,血管支架植入术39例,急诊溶栓术27例）介入治疗术后即刻CT图像表现。根据其位置、CT值、高密度量及24 h随访结果,术后CT图像颅内高密度病灶被分为3型,即Ⅰ型为正常对比剂分布,Ⅱ型为对比剂滞留或渗出,Ⅲ型为颅内出血。结果：神经介入术后即刻头颅CT检查示Ⅰ型74例（71.8%）;Ⅱ型25例（24.3%）,其中颅内动脉瘤栓塞术8例,支架植入术6例,动脉溶栓11例;Ⅲ型4例（3.9%）,其中颅内动脉瘤栓塞术1例,动脉溶栓3例。结论：正确诊断和鉴别神经介入术后即刻头颅CT高密度表现对指导术后治疗有重要价值。     

The DePICTORS Study: discrepancies in preliminary interpretation of CT scans between on-call residents and staff

At many academic hospitals, radiology residents provide preliminary interpretations of CT studies performed outside of regular working hours. We examined the rate of discrepancies between resident interpretations and final reports issued by staff. We prospectively obtained 1,756 preliminary reports and corresponding final reports for computed tomography (CT) scans performed on call between November 2006 and March 2007. The overall rate of clinically significant discrepancies (those that would potentially alter the patient’s clinical course prior to issue of the final report) was 2.0%. Major discrepancy rates for abdominal/pelvic, chest, cervical spine and head CT were 4.1%, 2.5%, 1.0% and 0.7%, respectively. Senior residents had fewer major discrepancies compared to their junior colleagues. Time of interpretation was also evaluated, but a statistically significant relationship was not observed. In summary, this study demonstrates a low discrepancy rate between residents and staff radiologists and identifies areas where after-hours service may be further improved.     

婴儿期迟发性维生素K缺乏致颅内出血的CT表现

目的探讨婴儿期迟发性维生素K缺乏性颅内出血的CT影像特点。方法选择临床确诊为迟发性维生素K缺乏性颅内出血的23例患儿的头颅CT资料进行回顾性分析。结果蛛网膜下腔出血8例,脑实质出血10例,硬膜下出血5例,脑室内出血3例,其中60%为多部位出血,脑梗死1例。结论结合临床特点,CT检查不仅可以明确具体的出血部位,还可以辅助作出出血的病因诊断。     

Uncommon locations of hydatid disease: CT appearances

Hydatid disease (HD), already known by Hippocrates, is prevalent and widespread in most sheep-raising countries in Asia, Australia, South America, Near East, and southern Europe. The disease is most commonly due to Echinococcus granulosus and may occur in any organ or tissue. The location is mostly hepatic (75 %) and pulmonary (15 %), and only 10 % occur in the rest of the body [1]. Imaging modalities such as US, CT, and MR imaging are helpful in diagnosing the disease. The reliability of each method depends on the cyst's location in the body [2–5]. The purpose of this essay is to illustrate the use of CT in depicting some unusual locations of HD besides the liver and lung. Received 4 April 1996; Revision received 26 August 1996; Accepted 8 November 1996     

Dual-energy bone removal computed tomography (BRCT): preliminary report of efficacy of acute intracranial hemorrhage detection

Purpose

One of the major applications of dual-energy computed tomography (DECT) is automated bone removal (BR). We hypothesized that the visualization of acute intracranial hemorrhage could be improved on BRCT by removing bone as it has the highest density tissue in the head. This preliminary study evaluated the efficacy of a DE BR algorithm for the head CT of trauma patients.

Methods

Sixteen patients with acute intracranial hemorrhage within 1 day after head trauma were enrolled in this study. All CT examinations were performed on a dual-source dual-energy CT scanner. BRCT images were generated using the Bone Removal Application. Simulated standard CT and BRCT images were visually reviewed in terms of detectability (presence or absence) of acute hemorrhagic lesions.

Results

DECT depicted 28 epidural/subdural hemorrhages, 17 contusional hemorrhages, and 7 subarachnoid hemorrhages. In detecting epidural/subdural hemorrhage, BRCT [28/28 (100%)] was significantly superior to simulated standard CT [17/28 (61%)] (p = .001). In detecting contusional hemorrhage, BRCT [17/17 (100%)] was also significantly superior to simulated standard CT [11/17 (65%)] (p = .0092).

Conclusion

BRCT was superior to simulated standard CT in detecting acute intracranial hemorrhage. BRCT could improve the detection of small intracranial hemorrhages, particularly those adjacent to bone, by removing bone that can interfere with the visualization of small acute hemorrhage. In an emergency such as head trauma, BRCT can be used as support imaging in combination with simulated standard CT and bone scale CT, although BRCT cannot replace a simulated standard CT.