| BackgroundIntracerebral hemorrhage is the most common type of stroke,accounting for 15%-20%of all strokes,with an annual incidence of 24.6/100,000.Meanwhile,intracerebral hemorrhage has the worst prognosis among all types,with a 1-month mortality of 40.4%and a 1-year mortality of 54.7%.The disability rate of surviving patients is as high as 75%,and there is currently a lack of effective treatment.Hospital admissions for cerebral hemorrhage,now one of the world’s most serIOUs public health problems,have increased gradually in the past decade.Hematoma enlargement is the only potential therapeutic target that is highly correlated with the prognosis of intracerebral hemorrhage and has the possibility of intervention.Actively exploring the clinical indicators that can predict the occurrence of hematoma enlargement and looking for the therapeutic means that may affect the expansion of hematoma have become one of the hot spots of clinical research at present.Therefore,it is necessary to accurately evaluate the volume of hematoma to assist the formulation of diagnosis and treatment plan and prognostic prediction.Machine learning has surpassed human in its ability to identify complex patterns using quantified image information,and has recently developed vigorously in the application of medical images.Deep learning technology represented by deep convolutional neural network has shown great potential in assisting clinical decision making.ObjectiveThe relative generalized deep learning algorithm was applied to segment and accurately evaluate the hematoma volume to reduce the error of artificial calculation,and the CT signs related to the risk of early expansion of cerebral hematoma were recognized by artificial intelligence,and the deep learning prediction model of early expansion and poor prognosis of cerebral hematoma is built on the basis of the above research.It aims to to explore the diagnostic efficacy of deep learning in early extension and poor prognosis of supratentorial spontaneous intracerebral hemorrhage,and to provide a new computer-aided diagnostic tool for clinic.Materials and methodsCT images and clinical data of 451 cases of supratentorial spontaneous intracerebral hemorrhage from four hospitals were retrospectively collected.Firstly,MSHA-Unet deep learning algorithm model was constructed to segment cerebral hematoma,and the segmentation effect was compared with other five deep learning network models,and the volume measurement was compared with Tada formula.Then,based on CT images,the island signs,swirl signs,black hole signs,blend signs,and satellite signs were identified by Resnet network,and evaluate the recognition efficiency.Finally,the deep learning algorithm was used to establish the model,and compared with five machine learning algorithm prediction models,to compare the prediction performance of different learning models for early hematoma expansion and poor prognosis of patients with intracerebral hemorrhage,and to compare the advantages and disadvantages of each algorithm.Results1.Supratentorial hematoma MSHA-Unet with deep learning algorithm model:(1)Hematoma segmentation:The Dice value of MSHA-Unet model was 0.96,which was better than other network models and had statistical significance(P<0.05).There was no difference between DeepLabV3plus and MultiResUnet in Jaccard similarity coefficient,but there was difference among other groups.There was no difference in sensitivity Sen value among all groups.(2)Comparison with other hematoma volume calculation methods:The hematoma volume obtained by the model algorithm was closer to that obtained by manual segmentation.Compared with the Tada formula group,the percentage error of the model algorithm group was smaller(8.197%),and the intra-group correlation coefficient was 0.987.96%(48/50)of the data fell within the 95%consistency limit(LoA),and its 95%LoA was narrower.It was 1.70～4.06 mL.2.Artificial intelligence recognition of early dilatation-related CT signs of supratentorial hematoma:(1)Comparison and analysis of the dilated and non-dilated hematoma group:the probability of "black hole signs","mixed signs","island signs","vortex signs" and "satellite signs" on CT scanning in the dilated hematoma group was significantly higher than that in the non-dilated hematoma group(P<0.05).(2)Using Resnet deep neural network to build CT feature recognition model:Among the five CT signs associated with early expansion of hematoma,the recognition efficiency was significantly improved after accurate segmentation and processing of hematoma.Among them,the recognition efficiency of island sign and black hole sign was higher than other signs,with accuracy rates of 76.71%(95%CI:65.35%-85.81%)and 72.13%(95%CI:57.65%-83.21%),AUC was 0.746(95%CI:0.631-0.841)and 0.667(95%CI:0.525-0.791),and specificity was 81.63%(95%CI:68.58%-92.18%)and 89.74%(95%CI:74.97%98.02%),the sensitivity of the island by 14.81%(95%CI:4.19%33.73%)to 65.38%(95%CI:46.04%83.48%).3.Construction of deep learning prediction model for early expansion and poor prognosis of hematoma:(1)Prediction model for early expansion of hematoma:Only the imaging signs of cerebral hematoma were used to predict the expansion of cerebral hematoma.AUC of all models in the training set was 0.694-0.839,while AUC of all models in the test set was generally low,ranging from 0.597-0.691.The AUC value of the Simple Bayesian algorithm was the highest,and the AUC of the deep learning model in the test set was 0.623.Combined with imaging and clinical indicators,the AUC of the deep learning model in the test set was 0.705,slightly higher than that of most machine learning models.Random forest model was a stable model that performs well in both training set and test set.(2)Prediction model of poor prognosis of intracerebral hemorrhage:Compared the efficacy of prediction models for poor prognosis of patients with intracerebral hemorrhage,among the early prediction models using initial hematoma volume combined with relevant imaging and clinical indicators,the prediction performance of all models was high.In the training,AUC value reached 0.872-0.923.In the test set,the AUC value of each model decreased slightly,and the AUC value was 0.824-0.878.(3)Correction of prediction model for poor prognosis of intracerebral hemorrhage:The hematoma volume at the time of review was used to correct the prediction model,and the prediction performance of most models after correction was roughly similar to that of the early model.In the training set,the AUC was 0.89-0.919.The AUC of the test set was 0.803-0.888.The AUC of the deep learning model in the test set was 0.888.Conclusion1.MSHA-Unet network model can better solve the segmentation problem of cerebral hematoma,and it is a feasible way to reduce labor cost and build a more generalized capability model by using data sets to establish segmentation model.2.CT features related to early expansion of brain hematoma including island sign and black hole sign can be effectively recognized based on deep learning algorithm,while the recognition efficiency of vortex sign,satellite sign and mixed sign is limited,and the overall recognition accuracy needs to be improved.3.The algorithm model built based on deep learning can better predict the early expansion and poor prognosis of cerebral hematoma,so as to more comprehensively evaluate the situation of supratentorial spontaneous intracerebral hemorrhage and guide clinical diagnosis,treatment and prognosis prediction. |
PDF Full Text Request