A Study Of Automated ASPECTS In Acute Ischemic Stroke Based On Multimodal CT

Part one The Value of Lower-Dose Non-contrast CT in Automated ASPECTSBackground and purposeLower-dose scanning protocols have been widely applied in the clinical workflow of acute ischemic stroke(AIS).However,the value of lower-dose non-contrast CT(NCCT)in automated Alberta Stroke Program Early CT Score(ASPECTS)is still unclear.To evaluate the feasibility of lower-dose NCCT in automated ASPECTS and investigate the influence of slice thickness(ST)and the reconstruction algorithm on the agreement of the automatically calculated scores,in order to select the best ST and reconstruction algorithm for automated ASPECTS.Materials and methodsFor this retrospective study,reference standard was defined by three radiologists based on baseline NCCT collected from March 2018 to June 2020 from patients who received intravenous thrombolysis(IVT)or endovascular treatment(EVT).Imaging data from both baseline and follow-up imaging was evaluated for the reference standard.Baseline NCCT was acquired with a lower-than-standard-dose imaging protocol(120 KV,200 m As)using hybrid iterative reconstruction(HIR)with 5 mm ST and 1 mm ST,as well as iterative model reconstruction(IMR)with 1 mm ST.From the baseline NCCT,baseline ASPECTS were generated automatically by two software packages.Follow-up ASPECTS was calculated by one radiologist based on the follow-up imaging.The intra-class correlation coefficients(ICC)were used to examine the agreement between baseline ASPECTS and reference standard.Accuracy,sensitivity,specificity,and area under the curve(AUC)of the receiver-operating characteristics(ROC)analysis were used to assess the performance of automated ASPECTS compared to reference standard at the region level.The correlations of baseline ASPECTS with baseline stroke severity(NIHSS)and follow-up ASPECTS were evaluated withSpearman correlation analysis.ResultsTwo hundred seventy-six patients met our study criteria.In score-based analysis,RAPID software showed almost good agreement with reference standard with ICC ? 0.74 and NBC software showed moderate agreement with reference standard with ICC ? 0.59 for any CT reconstructions.In region-based analysis,sensitivity of RAPID software ? 0.54,specificity ? 0.91,accuracy ? 0.83 and AUC ? 0.74.Sensitivity of NBC software ? 0.51,specificity ? 0.81,accuracy ? 0.73 and AUC ? 0.66.The sensitivity and AUC of baseline ASPECTS calculated by the two software packages(RAPID: 0.60,0.76;NBC: 0.56,0.71)using 5 mm HIR were superior to those calculated using 1 mm HIR(RAPID: 0.54,0.74;NBC: 0.51,0.66),with statistical significance(P < 0.05).The AUCs of baseline ASPECTS calculated by the two software packages(RAPID: 0.78;NBC: 0.72)using 1 mm IMR were also superior to those calculated using 1 mm HIR(RAPID: 0.74;NBC: 0.66),with statistical significance(P < 0.05).Baseline ASPECTS were significantly associated with baseline NIHSS and follow-up ASPECTS.ConclusionsLower-dose NCCT is feasible in automated ASPECTS.At score level,automated ASPECTS can reflect the size of early ischemic changes(EIC)in AIS.At region level,automated ASPECTS varied in different brain regions,thereby the affected region in AIS need to be identified by radiologists.For lower-dose NCCT with different CT reconstructions,5 mm was the optimal ST.IMR improves the reliability of automated ASPECTS.Part two Establishment and evaluation of an artificial intelligence model in ASPECTS assessment of acute ischemic stroke based on NCCTBackground and purposeThe distribution of EIC in the MCA territory according to ASPECTS is related to functional prognosis of AIS.However,affected by the accuracy of region segmentation,region-based automated ASPECTS results varied in different brain regions.The aim of this study was to develop and validate an automatic ASPECTS model that utilizes a three- dimensional convolutional neural network(3D-CNN)segamentation model and a machine learning model based on radiomics to differentiate between normal and EIC.Materials and methodsWe retrospectively collected 513 AIS patients(1026 brain regions)who underwent multimodal CT at baseline from July 2016 to July 2020.According to the stratified sampling method,100 patients were selected as the test group,and the remaining 413 patients were used as training/validation group.In order to ensure that the number of affected regions of ASPECTS,the number of patients with chronic ischemic stroke and leukoaraiosis in the test group and training/validation group are not less than 1:4.ASPECTS regions manually contoured on NCCT were used as the reference standard for segamentation model.The performance of 3D-CNN segamentation model was measured by dice similarity coefficient(DSC)and average hausdorff distance(AHD).On the basis of the segmentation model,for each region of ASPECTS,the radiomics features were extracted and selected.Applying 9 machine learning algorithms to develop classification model to discriminate between normal or EIC.The machine learning model was trained and validated on 413 patients using 5-fold cross-validation.The model was further tested on an independent test group of 100 patients.Furthermore,four radiologists,who were blinded to the reference standard and the artificial intelligence model's results,independently evaluated NCCT of the test group.The ICC were used to examine the agreement between automated ASPECTS,manual ASPECTS and reference standard.Accuracy,sensitivity,specificity,and area under the curve(AUC)of the receiver-operating characteristics(ROC)analysis were used to assess the performance of automated ASPECTS compared to reference standard at the region level.Results(1)The DSC values of region-specific segamentation model(C,IC,L,I,M1 – M6)were 64%,63%,76%,77%,80%,80%,75%,77%,79% and 77%.(2)The AHD values of region-specific segamentation model(C,IC,L,I,M1 – M6)were 1.24,0.86,0.94,1.16,1.24,1.26,1.37,1.33,1.50 and 1.55.(3)Accuracy,sensitivity specificity and AUC of the region-specific machine learning model(C,IC,L,I,M1 – M6)were ranged from 62%-90%,43%-91%,81%-96% and 0.72-0.93.(4)Automated ASPECTS from the artificial intelligence model agreed well with reference standard(ICC = 0.78).Automated ASPECTS was non-inferior to experienced radiologists in scoring ASPECTS on NCCT(P > 0.05).Automated ASPECTS was superior to junor radiologists(P < 0.05).(5)Region level analysis showed that the artificial intelligence model yielded accuracy of 80%,sensitivity of 76% and specificity of 82%.The model's sensitivity was superior to four radiologists(P <0.05),accuracy was superior to junor radiologists(P < 0.05)and noninferior to experienced radiologists(P > 0.05).ConclusionsThe 3D-CNN model provided a tool to realize automatic segmentation of ASPECTS region.It has good segmentation accuracy and high consistency with manual segmentation,which can be applied in automated ASPECTS assessment.The artificial intelligence model was established for the assessment of ASPECTS with favorable diagnostic performance,which can effectively assist in ASPECTS assessment and is expected to improve the efficiency of diagnosis.Part three Post-processing of computed tomography perfusion in patients with acute ischemic stroke: a reproducibility studyBackground and purposeComputed tomography perfusion(CTP)plays an important role in the setting of reference standard for ASPECTS.Multiple factors have been identified to influence the evaluation of quantitative CTP parameters.To determine the reproducibility of quantitative CTP parameters generated using different post-processing methods and identify the relative impact of subjective factors on the robustness of CTP parameters in AIS,and to provide a basis for the setting of reference standard for ASPECTS.Materials and MethodsA total of 80 CTP datasets from patients with AIS or transient ischemic attack(TIA)were retrospectively post-processed by two readers using different regions of interest(ROI)types,input models,and software.The CTP parameters were derived for 10 parenchymal ROIs.The ICC were used to assess the reproducibility of the CTP parameters for various post-processing methods.The Spearman correlation test was used to detect potential relationships between software and input models.ResultsThe ICCs with 95% CI were 0.94(0.93–0.96),0.94(0.92–0.96),0.82(0.79–0.86),and 0.87(0.85–0.90)for inter-reader agreement by using elliptic ROI,irregular ROI,single-input model,and dual-input model,respectively.The ICCs with 95% CI were 0.98(0.98–0.98),0.46(0.43–0.50),and 0.25(0.20–0.30)for inter-ROI type,inter-input model,and inter-software agreement,respectively.The Spearman correlation coefficients between different input models based on cerebral blood volume(CBV),cerebral blood flow(CBF),mean transit time(MTT),and time to peak(TTP)were 0.60,0.79,0.56,and 0.50,respectively.The Spearman correlation coefficients between different software were 0.52,076,0.51,and 0.72 for CBV,CBF,MTT,and TTP,respectively.ConclusionAlthough the CTP parameters were stable when measured using different readers with different ROI types,they varied for different input models and software.The standardization of CTP post-processing is essential to reduce variability of CTP values and to provide a basis for the setting of reference standard for ASPECTS.