Research On Deep Learning Method Of Small Sample Weak Label Medical Images

Medical imaging-assisted detection technology is one of the important technologies for cancer diagnosis.The analysis of medical image data based on deep learning technology can detect the distribution of pathological information and the nature of the lesion area,so as to give a reasonable diagnosis method and the future development trend of patients.Judging whether the target has hidden cancer risks is of great research significance based on the screening results And research value.In this paper,the improved deep learning method is used to detect the diseased area in the breast medical image.In view of the current detection algorithm's problems of high false detection rate and high miss rate of small units,this paper proposes a breast cancer detection method based on improved convolutional neural network based on improved feature pyramid structure.This method combines the computerized tomography(CT)medical image data set of breast to analyze the annotation information and structural characteristics of the lesion area in the data.On this basis,the residual network is used as the feature extraction network to extract the multi-scale features of the detected image,Using the feature pyramid structure of the front end of Mask R-CNN,on the basis of top-down multi-scale feature fusion,horizontally add bottom-up feature fusion,combined with multi-scale high-level semantic features and low-level texture features,to improve the accuracy of breast cancer detection rate.Experimental results show that under the same feature extraction network and data set,the detection accuracy and speed of this method are improved to some extent compared with other detection algorithms.Aiming at the problems of low data volume and low sample detection accuracy caused by the small sample weak label medical image data set,this paper proposes a detection method based on difficult sample mining.This method analyzes the data characteristics of small-scale magnetic resonance imaging(Magnetic Resonance Imaging,MRI),obtains a pre-trained model in the CT image data domain with similar characteristics,uses secondary transfer learning to continue training,and combines the source domain and the target domain.Relevant feature information is helpful to solve the problem of small samples.Aiming at the imbalance of positive and negative sample ratios,a hard sample mining mechanism is introduced to screen the hard negative samples generated during the calculation of the classification loss of the model,and the network model is used to further learn the feature information of some samples to improve the classification accuracy of the model.The experimental results show that,compared with Mask R-CNN and one-stage detection model,the method in this paper can detect diseased areas in MRI test data with high accuracy,and the detected diseased areas in breast images have high classification Accuracy.