Purpose:MRI is an important method to diagnose cervical spine diseases in clinical practice.In addition,deep learning technology has also good recognition results in the field of medical image.This paper uses deep learning technology to identify and detect the lesions of MRI images for cervical spinal cord injury and disc disease,and provides a feasibility study for its clinical application.Materials and Methods:(1)Data set: patients with cervical diseases were admitted to our hospital from January 2013 to December 2018.We defined two kinds of diseases as inclusion criteria: cervical disc degenerative diseases(DDD)were mainly refers to cervical disc herniation and traumatic spinal cord injury(SCI)were mainly refers to spinal cord signal changed due to injury,at the same time,spinal cord tumors,syringomyelia,motor neuron disease(MND),peripheral polyneuritis as exclusion criteria.Total 1,010 patients were enrolled from the picture archiving and communication systems(PACS)station,including 692 male patients,318 female.In addition,another 500 patients were collected with diagnosing negative(patients without DDD and SCI)in order to get better training results in reality.Each patient’s images were respectively divided into the following categories: ‘normal group’,‘disc group’ and ‘injure group’.At last,all images were desensitized before using(e.g.,patient name,age,date of examination,etc.).(2)Grouping of data: to simulate the proportion of the incidence in reality.Specifically,the data set is divided into three groups: "normal group"," disc group",and "signal change group".This data set was randomly split into two parts: 1210 patients(approximately 80%)for training and 300 patients(approximately20%)for validation.(3)Labeling and preparation: the data set is processed before the experiment.In order to acquire a higher prediction effect,the training set and validation set usually used bounding box to annotate the lesion’s area,while ‘normal group’ without bounding box.In this process,the bounding box were labeled by two experienced spine surgeon using Label Me Tool box-master.(4)Methods: we designed the relevant parameters and performed Faster-region convolutional neural networks(Faster R-CNN)combined with a backbone convolutional feature extractor using Res Net-50 and VGG-16 network to detect the lesion on MRI owing to cervical diseases.(5)Testing set: in order to optimize the ratio reasonably,additional 500 MRI images as testing set to demonstrate detection performance of our method.The number of "normal group"," disc group",and "signal change group" is 200,200 and 100,respectively.(6)Result evaluation: the mean average prediction accuracy(mAP)and visualization results as the evaluation to measure the prediction effect of these two methods.Result:After testing,the results showed that Faster R-CNN with Res Net-50 and VGG-16 had good prediction and detecting speed for detection and recognition the lesions from cervical spinal cord MRI.The mean average precision(mAP)was 88.6%,72.3%,respectively,and testing time speed was 2.2,2.4 sec/image,respectively.In addition,visualizing data and numerical coordinates are also showed in results.In terms of prediction effect and detection time,Faster R-CNN algorithm with Resnet50 is better than Faster R-CNN algorithm with VGG-16.Conclusion:(1)Our neural network model using Faster-RCNN algorithm based VGG-16 and Res Net50 as backbone could be used to identify diseases on MRI in spinal cord.(2)Using standardized data sets,reasonable model selection and application of evaluation standards can achieve the expected results of the experiment.Compared with the VGG-16 model,the Res Net50 model has better in prediction results,faster detection speed,its showed that indirect the deeper network structure has better effect on prediction,and its also confirms that the effectiveness of the model of big data + artificial intelligence.(3)It might provide a possibility of aid-diagnosing for radiologists and spine surgeons.This evidence motivates future research into better combining the clinical and medical imaging basis of deep learning networks as well as test in prospective data. |