Detection Of Cervical Spine Disease With Small MRI Dataset Based On SEnet

With the change of lifestyle,the number of long-term desk staff has increased.In recent years,the incidence of cervical spondylosis has continued to rise,and there has been a shortage of clinical and radiologists.In clinical practice,doctors are prone to miss diagnosis and misdiagnosis of cervical spinal cord disease.The application of deep learning methods in the medical field is currently called the focus of attention.Among them,a target detection algorithm using a convolutional neural network can quickly detect the position of a diseased area in a medical image and classify a disease type.If deep learning methods can be applied to the diagnosis of cervical spinal cord disease,it will not only speed up the diagnosis of the disease,earn valuable time for the treatment of patients,but also reduce the rate of misdiagnosis and missed diagnosis.In this paper,both cervical MRI images and convolutional neural network target detection algorithms are studied in depth.At present,there is no publicly available cervical MRI image dataset for research in academia.Therefore,the author first established a small cervical MRI dataset in collaboration with doctors of the First Medical College of Bethune,Jilin University,and proved the validity of the dataset through the mature natural image target detection algorithm faster-rcnn and the use of deep learning to detect cervical spinal cord disease.Feasibility.After that,an improvement method is proposed for the over-fitting problem of the faster-rcnn algorithm on this data set due to the small size of the data set.First,the elastic deformation algorithm is used to simulate the deformation inherent in the patient's MRI images,and to enhance the data set.Second,adding the SE module to the feature extraction module of the algorithm to adjust the network structure to the SEnet structure.This change improves the accuracy of the network based on data enhancement.Third,the gated layer is used to integrate three weighted images commonly used in cervical MRI,and other methods that combine different inputs are discussed.Through the above methods,the improved target detection algorithm proposed in this paper achieves a precision of 87.53% and a mAP of 0.76 on a small sample of cervical MRI dataset at 95% recall.Higher than faster-rcnn method and radiologist level.