Precise Diagnosis Of Intracranial Hemorrhage And Subtypes Using A Deep Learning Algorithm Model With CNN And RNN

Objectives To reduce the misdiagnosis rate of radiologists in the diagnosis of intracranial hemorrhage to a certain extent,we use the deep learning method to construct a new 3D neural network model of convolutional neural network combined with recurrent neural network(CNN-RNN)and evaluate it.The value of intracranial hemorrhage(ICH)and its five subtypes(cerebral parenchymal,intraventricular,subdural,epidural and subarachnoid)in non-contrast head CT.Methods 2,836 subjects(ICH/normal: 1,836/1,000)from three institutions were included in this ethically approved retrospective study,with a total of 76,621 slices from non-contrast head CT scans.Maintain a high ICH ratio(65%)in the data set to ensure that there are enough positive samples to aid in the learning process of the algorithm and to effectively evaluate the performance of the algorithm.In order to provide trainable data,the original CT image is preprocessed using the following steps.First,all image slices are resampled to 512 x 512 pixels and then downsampled to 256 x 256 pixels to reduce GPU memory usage.Considering that different CT window widths and window levels may be required for CT image viewing,and the differences in details of different tissues are preserved,we have selected three sets of windows with different CT values(Hounsfield units,HU)to normalize the image:-50-150 HU,100-300 HU and 250-450 HU.Each slice of ICH and its five subtypes were independently performed by three senior radiologists,and the results of the three final consultations were used as reference standards for case-level and slice-level image labels.80% of the data is used for model training,10% of the data is used for model validation,and the remaining 10% is used for final testing of the model.We propose a joint CNN-RNN classification framework that provides the flexibility to train case-level or slice-level labels.The algorithm model test results were compared with those of three radiologists standardized training physicians and another senior radiologist.In addition,in order to improve the interpretability of the model,we use the Grad-CAM method to generate a rough positioning map in the algorithm,highlighting the anomalous area in the image.The positioning map on each slice is generated using the training algorithm,which does not affect the algorithm training process,nor does it need to manually delineate the bleeding area for supervised training.This visualization technique can also be used by radiologists as a basis for retrospective review.To some extent,it solves the lack of transparency and interpretability of the deep learning model.Results For the 2-class task(predicting the presence of intracranial hemorrhage),the accuracy of the algorithm at the case level was high(≥0.98).For the 5-class task(predicting five intracranial hemorrhagic subtypes),the algorithm had an AUC and specificity of > 0.8 on all subtypes,and the SAH and EDH subtypes were significantly less sensitive than the other three types.Using the Grad-CAM method to generate a heat map,you can initially locate the anomaly area and visualize the results.For the 2 and 5 classification tasks,the algorithm does not perform as well as the senior radiologist compared with the diagnostic results of the radiologist,but the algorithm can correct the misdiagnosed cases to some extent,and the algorithm is compared with the standardized training trainers of the radiology department.The average performance of the standardized trainees was better when predicting intracranial hemorrhage.Conclusions The algorithm proposed in the study is fast and accurate,which helps to help the radiologists with insufficient head CT experience to reduce the rate of misdiagnosis of the initial diagnosis.In clinical practice,the application of this automated classification system helps to improve the efficiency of radiologists and reduce work stress.