Research Of Intelligent Aided Diagnostic Methods For Multi-modality Medical Imaging

With the rapid increase in the data amount and the continual modality updating of medical images,there are increasing demands for intelligent,automated,and highly accurate computeraided diagnosis(CAD)technology in medical community.Artificial intelligence(AI).the most advanced technology in the information field,will provide new hopes and ideas for precision medicine,and medical imaging aided diagnosis based on artificial intelligence is one of the earliest areas to achieve relevant breakthroughs and outbreaks.Therefore,based on the study of masses of domestic and foreign latest literatures,this paper combined with artificial intelligence,mathematics,image processing,medicine and other theories to research novel technologies and methods of intelligent diagnosis for multi-modality medical imaging according to the clinical demand of doctors,using digital pathology and PET/CT as the main experimental data.The main research results were as follows:(1)Digital pathology aided diagnosis for hepatocellular carcinoma(HCC)Liver cancer nest location and HCC nucleus detection on pathological images:To locate the cancer nest region on the whole pathological image,this paper proposed an inputcollaborative parallel‘same’ and‘valid’ convolutional residual neural network(IC-PSV-RNet).The‘same’convolution output the feature map with the same size of each input,and highlighted the local features.The ’valid’ convolution extracted the feature map by considering the downsampling simultaneously.This two added feature maps can highlight the region of interesting(ROI).The input-collaboration strategy can integrate global information of the original input and learn local features of the convolutional block(CB).Inspired by residual learning,we added the input of each CB into its output in feature dimension via the identity mapping form,which can prevent degradation of network performance and strengthen the relationship between each layer of features.The experimental results of cancer nest location demonstrated that IC-PSVRNet was superior to ResNet and DenseNet.In addition,this paper developed a domain adaptive nucleus detection model based on stacked sparse auto-encoder(SSAE),its advantage was to use a detection model trained by only one kind of pathological images to detect nuclei on other kinds of pathological images a accurately.The nucleus detection results on liver pathological images demonstrated that SSAE-based domain adaptive model can perform online learning and was superior to other CNN-based,DBN-based detection models.HCC nucleus segmentation on pathological images:In order to segment adhesion and overlapping nuclei on cancer nest,this paper proposed a coarse-to-fine segmentation framework based on SC-ELM and CBST.For the coarse segmentation:SC-ELM provided a novel theory for structural classifier,which extracted and refined image features via random kernels and only learned the output weight matrix to predict the segmentation result with the same resolution of each input.For the fine segmentation:we constructed the corresponding CBST via each patient’s coarse segmentation result,and then estimated the lost nucleus boundary via the pixelwise classification.The experimental results demonstrated that the Dice coefficient based on SC-ELM can achieve 0.959.HCC multi-classification and HCC nucleus grading on pathological images:To perform pathological image multi-classification,this paper proposed an input-collaborative parallel‘same’ and ’valid’ convolutional neural network(IC-PSV-Net)in an end-to-end form.The input-collaboration strategy referred to transfer grayscale information of the original input to the feature maps of each CB,which can compensate the feature attenuation and further improve classification performance.The experimental results on lymphoma,breast cancer and HCC pathological image dataset demonstrated that IC-PSV-Net was superior to ResNet,DenseNet and other CNN models.Besides,this paper proposed a joint multiple fully connected convolutional neural network with extreme learning machine(MFC-CNN-ELM)for HCC nuclei grading.The core of MFC feature optimization model was to use different strides and sizes of filters to further excavate discriminative features,which can improve the performance of shallow layer’s CNNs.The experimental results on HCC nucleus grading demonstrated MFC-CNN-ELM achieved the accuracy of 0.967 and was superior to other traditional CNNs with single full connection layer.(2)Radiomics aided diagnosis for lymphomaMulti-organ segmentation on CT images:To reduce the false positive rate of lymphoma detection and determine lymphoma stage,inspired by U-net,IC-PSV-SegNet,an end-to-end,was developed for segmenting multi-organ on CT images,which added the up-sampling layer into IC-PSV-Net and introduced the mechanism of feature connection.Then,we proposed a contour sparse representation(CSR)model to optimize the organ’s local boundary,its core was to first regard a closed curve as a set of continuous points,then establish the model of finding boundary feature points via the posterior probability model and solve it by introducing gradient,rate change,probability regularization and appearance energy functions,finally optimize boundary via the pixel-wise classification.The experimental results on multi-organ segmentation demonstrated the proposed method achieved Dice coefficients of 0.945,0.935 and 0.973 on liver,kidney and spleen respectively,and was superior to traditional 3D U-Net and 3D FCN.Lymphoma detection and segmentation:For lymphoma detection,this paper used ICPSV-Net to identify whether each possible candidate was a real lymphoma,where the inputcollaboration strategy was to transfer the gradient information of original CT into the feature map of each CB for improving detection accuracy by combining PET metabolic and CT structural information.Then,we proposed an adaptive weighting and generalized distance regularized level set evolution(AW-GDRLSE)for delineating lymphoma accurately,which can provide the automatic annotation for surgical navigation and radiation therapy.AW-GDRLSE improved the traditional level set method from three aspects:① Introducing a generalized distance regularization to maintain the contour’s regular shape during the evolution process and using the parameter q to control the contour’s diffusion rate;② Defining the local regionbased energy function via an annular mask;③Calculating weights of the length and area term adaptively via local region intensity and boundary direction information in each evolution process.The experimental results on lymphoma segmentation demonstrated that AW-GDRLSE was superior to traditional C-V,DRLSE and other level set models in terms of segmenting small tumor with the Dice coefficient of 0.8277.