Application Of Artificial Intelligence In The Early Diagnosis And Prognosis Prediction Of Acute Ischemic Stroke

Objective: To develop artificial intelligence(AI)models for early diagnosis and long-term functional prognosis of patients with acute ischemic stroke(AIS)by combining novel AI technologies with routine imaging examinations used clinically to assess AIS patients,to assist radiologists make more accurate diagnoses and personalise patient prognosis,and to provide objective imaging supports for clinicians to make appropriate treatment decisions,allocate time and resources appropriately,and optimise rehabilitation programmes with realistic goals.Methods: Routine examination images of AIS patients and the control group,including head non-contrast computed tomography(NCCT),head magnetic resonance imaging(MRI),and diffusion weighted imaging(DWI),were retrospectively collected.Patients were grouped according to the follow-up images,clinical diagnosis,and assessment scales.The AI models were trained according to the images and features extracted from the images to identify the negative AIS lesions in the NCCT and to predict the long-term functional prognosis of patients.The diagnostic and predictive efficacy of the models were evaluated by the AUC,sensitivity,specificity,and accuracy values with 95% confidence interval.The Delong test was used to compare the differences in AUC values,and the Chi-square test was used to evaluate the differences in rates.Results: The AUC of the developed deep learning model was much higher than that of two experienced radiologists for detecting the negative AIS lesions in the NCCT,and both radiologists had improved accuracy in diagnosing AIS patients with the assistance of the model.The combined model based on the fusion of deep features extracted from the NCCT of AIS patients with clinical information had a better predictive performance than other models.The built fusion model combining lesion spatial information features and clinical information had good predictive efficacy for functional outcomes of AIS patients at 90 days.Conclusions: This study developed novel models for the early diagnosis and prognosis prediction of AIS from the routine imaging examinations used clinically to assess AIS patients,combined with new AI techniques,to bridge the gap between conventional neuroimaging in the early diagnosis and prognosis prediction of AIS.Part Ⅰ: Detection of negative acute ischemic stroke in non-contrast computed tomography using two-stage deep learning modelObjective: Although NCCT is the recommended examination for all patients with suspected AIS by international guidelines,it cannot detect significant changes even for experienced radiologists in early infarction.We aimed to develop a deep learning model to detect the negative AIS lesions in the NCCT images and evaluate its diagnostic performance and capacity for assisting radiologists in decision-making.Methods: In this multi-center,multi-manufacturer retrospective study,1136 patients with suspected AIS but negative lesions in NCCT were collected from two geographically distant institutions between May 2012 to May 2021.The AIS lesions were confirmed based on the follow-up DWI and clinical diagnosis.The deep learning model was comprised of two deep convolutional neural networks to locate and classify.The performance of the model and radiologists was evaluated by the AUC,sensitivity,specificity,and accuracy values with 95%confidence intervals.The Delong test was used to compare the AUC values,and the Chisquared test was used to evaluate the rate differences.Results: 986 patients(728 AIS,median age,55 years,interquartile range [IQR]: 47-65 years;664 males)were assigned to the training and internal validation cohorts.150 patients(74 AIS,median age,63 years,IQR: 53-75 years;100 males)were included as an external validation cohort.The AUCs of the model were 0.836(sensitivity,68.99%;specificity,98.22%;and accuracy,89.87%)and 0.763(sensitivity,62.99%;specificity,89.65%;and accuracy,88.61%)for the internal and external validation cohorts based on the slices.The AUC of the model was much higher than that of two experienced radiologists(0.655 and 0.595 in the internal validation cohort;0.640 and 0.644 in the external validation cohort;all P < 0.001).The accuracy of the two radiologists increased from 62.00% and 58.67% to 92.00% and 84.67% when assisted by the model for patients in the external validation cohort.Conclusions: This deep learning model represents a breakthrough in solving the challenge that early invisible AIS lesions cannot be detected by NCCT.The model we developed in this study can screen early AIS and save more time.The radiologists assisted with the model can provide more effective guidance in making patients’ treatment plans in the clinic.Part Ⅱ: Non-contrast CT-based deep learning model for prediction of long-term functional outcome in acute ischemic strokeObjective: To develop a deep learning(DL)model to extract deep features from NCCT images for predicting poor long-term functional outcomes in AIS patients.In order to assess the value of deep features in predicting prognosis,the extracted deep features were combined with clinical and imaging information to build multiple machine learning models for comparison of predictive performance.Methods: 485 patients with clinically confirmed AIS were retrospectively collected.The long-term functional outcome of patients was evaluated using the modified Rankin Scale Score at 90 days(m RS-90).Patients were divided into a good outcome group(m RS-90 ≤ 2)and a poor outcome group(m RS-90 > 2),and in a 7:3 ratio into the training and validation cohorts.A novel DL model(Model 1)was constructed and trained to extract deep features associated with poor outcome in the NCCT images and quantify them as DL score;Model 2 was built based on the patient modified Alberta Stroke Program Early CT Score(ASPECTS);Model 3 was built by combining clinical and imaging features;and Model 4 was built by fusing clinical,imaging,and DL features.The predictive performance of the four models was assessed by AUC,accuracy,precision,sensitivity,and specificity with 95% confidence intervals.The Delong test was used to compare differences in AUC values between models to assess the role of deep features in predicting poor outcome in AIS.Results: 485 patients with a mean age of 59.63±11.59 years,353(72.78%)had a good outcome and 132(27.22%)had a poor outcome.Four models were constructed to predict the poor outcome of AIS patients.In the training cohort,the predictive efficacy of Model 4(AUC = 0.928)was better than the other three models(P = 0.002,P < 0.001,P < 0.001);the predictive efficacy of Model 1(AUC = 0.883)was better than Model 2(AUC = 0.698,P < 0.001)and not significantly different compared to Model 3(AUC = 0.869,P = 0.602).In the validation cohort,Model 4 continued to have the highest predictive ability(AUC = 0.902);the predictive ability of Model 1 was consistent with its performance in the training cohort.DL score was correlated with independent predictors of clinical and imaging features related to prognosis,with the most positive correlation of age(Spearman’s correlation coefficient of 0.478,P < 0.001).The Nomogram constructed according to Model 4 had a high reliability.Conclusions: Deep features extracted from NCCT images using the DL model can make predictions about poor long-term functional prognosis in AIS patients.With the assistance of this model,clinicians can provide precise medical and personalized care for AIS patients.Part ⅡI: Prediction of long-term functional prognosis using a machine learning model with lesion spatial information in acute ischemic strokeObjective: To develop a machine learning model with lesion spatial information(LSI)to predict the long-term functional prognosis of AIS patients,and to compare the performance of the model with the DWI radiomics(DR)model and clinical models to explore the value of the LSI for prognosis.Methods: In this study,295 patients with AIS diagnosed from August 2018 to August 2020 were retrospectively collected as dataset 1.The modified Rankin Scale score at 90 days(m RS-90)was used to evaluate patients’ long-term functional prognosis.Included patients were divided into a good prognosis group(m RS-90 ≤ 2)and a poor prognosis group(m RS-90 > 2),with each group divided into the training cohort and the internal validation cohort in a ratio of 7:3.In addition,81 AIS patients were prospectively collected from September 2020 to December 2021 as an independent external dataset 2.The spatial information features and radiomics features of lesions were extracted from patients’ DWI images to bulid the LSI model and DR model,and the features were finally quantified into the LSI score and DR score.Patients’ clinical characteristics were collected to build a clinical model;LSI score,DR score and clinical characteristics were fused to build a fusion model.The predictive performance of each model was assessed by the AUC,accuracy,precision,sensitivity,and specificity values with 95% confidence intervals.The Delong test was used to compare the differences in AUC values between models to evaluate the role of LSI in predicting the prognosis of AIS.Results: The 295 patients in dataset 1,with a mean age of 58.65 ± 11.49 years,had a good prognosis of 207(70.17%)and a poor prognosis of 88(29.83%).The 81 patients in dataset 2 had a mean age of 56.62 ± 10.82 years,57(70.37%)had a good prognosis and 24(29.62%)had a poor prognosis.From 205 LSI features,6 white matter tract disconnection severity features and 1 grey matter parcel lesion load feature were screened for the LSI models,and 13 DWI radiomics features from 1210 were screened for the DR models.Among the clinical features of patients,age,diabetes,and baseline NIHSS score were independent predictors of poor prognosis.The predictive efficacy AUC of the fusion model was better than other models in different degrees,with the AUC of the training cohort,internal validation cohort and external validation cohort were 0.887,0.878 and 0.788,respectively.The AUC of the LSI model was higher than that of the DR model in the training cohort,internal verification cohort,and external verification cohort,but the difference was not statistically significant(P = 0.718,P = 0.773,P = 0.258).Conclusions: The integration of lesion spatial information in the prognostic model can improve the efficacy;the fusion model including the lesion spatial information and clinical information had a good predictive effect on the functional outcome of AIS patients at 90 days,which can provide a valuable reference for clinical treatment decisions of AIS patients.