Research About Deep Learning Based Two Stage Disease Diagnosis Method For Medical Image

During the era of Wise Medical(WIM),owing to the better basic medical facilities and explosive growth clinical data,diagnostic requirements of patients are increasing gradually.Traditional manual diagnosis methods dependent on clinicians and semiautomatic Computer-aided Diagnosis(CAD)based approaches have been unable to satisfy the needs of development.Purely automatic intelligent diagnosis proposals become the preferred.However,automatic disease diagnosis completely independent on doctors,especially ones relying on medical image analysis,is still a very challenging task.Since it is difficult to fully handle with each component of diagnosing processing,and previous CAD system does not involve relatively fresh intelligent techniques in recent time.Targeting to these defects,this paper proposes a two-stage disease diagnosis method in medical imaging domain based on deep learning(DL)techniques for achieving fully automatic medical image based pathological determination.It consists of two DL models,namely Multi-task Feature Supplemental Lesion Segmentation Network with Attention Mechanism(AMTFSLSN)and Deep Learning related Double-path Network Diagnosis Processing Flow(DLDPPF).Among them,AMTFSLSN combines supervised attention mechanism using to fix the lesion location with multi-scale feature supplementation containing region filtering to mitigate the situation about tumor pixel misjudgment.While DLDPPF utilizes dual convolutional neural network(CNN)feature extractor with feature redundancy control and maximum relevance based multilevel feature selection to exploit more diversified and more useful pathological features for improving the diagnosis precision.The lesion segmented by AMTFSLSN will be submitted into DLDPPF for assisting to finish disease diagnosis.Moreover,the two mentioned models decrease the training overhead by transfer learning(TL),and employ multi-path network architecture to increase the abundant degree of extracted features,which provides the foundation of model parallelization.Our work uses two challenging medical image datasets to test AMTFSLSN and DLDPPF respectively,including colorectal adenocarcinoma computed tomography(CT)dataset and breast cancer magnetic resonance imaging(MRI)dataset.The experimental results indicate that both segmentation model and diagnosis model obtain excellent performance,and significantly outperform compared state-of-the-arts,which proves the effectiveness and reliability of proposed two-stage medical image disease diagnosis method.