Research On Deep Learning Based Medical Image Classification

Medical imaging technology provides critical information for clinical diagnosis and treatment of a large number of diseases.In practice,the interpretation of medical image data is usually performed manually by experienced doctors.The process is tedious,timeconsuming,and heavily dependent on the subjective experience of experts.There is also a problem of judgment consistency among different experts.The introduction of computeraided diagnosis technology can greatly improve the efficiency of doctors.Existing research based on artificial features and traditional machine learning methods usually requires strong medical prior knowledge and tedious feature engineering,which limits the upper limit of their performance.Deep learning provides an end-to-end solution that can spontaneously learn the multi-level abstract features with good discrimination and generalization capabilities from the data,which significantly improves the accuracy of prediction results,so it quickly dominates various fields.This thesis studies medical image classification.Based on three different medical image datasets,including electroencephalography(EEG),magnetic resonance imaging(MRI),and time-lapse video,we study different medical diagnostic problems,and demonstrate different deep learning techniques and usages.Through various degrees of improvement and innovation in deep learning methods,our proposed approaches have reached the industryleading level in the corresponding medical research problems.We hope to provide a comprehensive perspective for the application of deep learning in medical imaging data.The main research contents of this thesis are as follows:(1)EEG and deep learning based sleep stage classification.Scoring of sleep stages plays an important role in the diagnosis of sleep-related diseases.At present,most of the existing deep learning methods are based on the original EEG signals,which requires a large amount of calculation and the algorithm performance is poor.In this chapter,we convert the single-channel EEG signal into a time-frequency map by short-time Fourier transform,and obtain inspiration from natural language processing.Two novel deep learning approaches are proposed for the detection of sleep stages.The C-CNN model based on convolutional neural network(CNN)is compact and efficient,achieving a good balance between computational cost and model performance;the Attention model based on attention mechanism and bidirectional long-term memory can get better performance.In addition,we integrate cost-sensitive learning into the model training process,which solves the serious class imbalance problem in sleep stages,ensures that each sleep stage has a higher recall rate,and achieves a better balance classification accuracy.(2)MRI and CNN based nasopharyngeal carcinoma diagnosis support system.Nasopharyngeal carcinoma is common in China and other Southeast Asian regions.This chapter proposes a visualization and decision-support system for the diagnosis of nasopharyngeal carcinoma,which can be run under different resolutions and MRI equipment.It first performs adaptive segmentation and cropping of MRI slices to extract the effective brain parts through the open operation and Otsu’s method,simultaneously solving the problem of different resolution and MRI equipment.Then,it uses the modified residual network to process MRI slices and extract features.Finally,the high-level abstract features of different slices are integrated to give the final positive probability of nasopharyngeal carcinoma.Our system can quickly visualize and locate the slices and regions where malignant tumors may exist,and the area under the ROC curve(AUC)for nasopharyngeal carcinoma positive diagnosis reached 0.994.(3)Multi-task deep learning with dynamic programming for embryo early development stage classification from time-lapse videos.During the treatment of in-vitro fertilization,accurate detection of the embryo early developmental stages can provide valuable information for the quality assessment of embryo,which is beneficial to the success of conception.This chapter proposes a multi-task deep learning framework with dynamic programming(MTDL-DP)for the classification of embryo early development stages.Firstly,based on the characteristic that adjacent frames in the video have lots of complementary information,it uses multi-task learning and adjacent frames to generate multiple predictions for each frame in the time-lapse video.Then,it integrates these predictions through ensemble learning to give the current frame one embryo development stage.Finally,dynamic programming post-processing is used to optimize the predicted embryo development stage sequence of the entire video,so that the final sequence is non-decreasing,and the loss of earth-mover distance is minimal.Through the proposed MTDL-DP approach,this chapter has improved the accuracy of embryo early development stage classification by 3.1%..