Deep Learning Stroke Prediction System For Multidimensional Segmentation Auxiliary Diagnosis And Multisource Feature Prognosis Evaluation

Background and ObjectiveStroke has the characteristics of high incidence,high recurrence rate,high disability rate and high mortality rate.Its diagnosis and prognosis evaluation are crucial and difficult aspects in clinical diagnosis and treatment.In view of the problems of limited research perspectives,single data sources,low method efficiency and lack of application promotion in previous studies,this paper concentrates on the relationship between image segmentation and radiomics to improve the effect of diagnosis and prognosis evaluation.In the image segmentation of auxiliary diagnosis,we investigate different dimensions of the segmentation algorithm and a variety of empirical data.The role of multiple methods and the inclusion of multisource feature sets will be investigated in radiomics-based prognostic assessment.The data and methods in this paper will help to enhance the diagnosis and prognosis of stroke and provide a decision-making basis for the diagnosis and treatment of stroke.Contents and MethodsThe primary line of the study is "image segmentation for auxiliary diagnosis →radiomics for prognosis evaluation".In auxiliary diagnosis,single-dimension and multidimension segmentation are conducted by scale bridging and dimension mixing,respectively.In prognostic assessment,improved prognostic assessment is conducted by feature fusion.(1)Scale bridging single-dimensional segmentation diagnosis prediction: This paper includes internal data,external data,multisource data and multimode data,constructs an edge enhanced skip connection(EESC)based on a scale bridging strategy,and then proposes a multiscale attention enhanced network(MAEN)to solve the problems of semantic scale bridging and edge feature loss in feature fusion of the encoder and decoder.(2)Multidimensional segmentation diagnosis prediction of dimensional mixture:Combining the advantages of 2D and 3D segmentation,this section constructs a multidimensional 2.5D segmentation model based on the improved strategy of dimensional mixing,i.e.,multidimensional hybrid evolutionary network(MHEN),and compares different methods such as full parallel,semiparallel and series.In the application promotion,the multidimensional segmentation interface is designed and implemented based on Py Qt.(3)Prognostic prediction of radiomics with feature fusion: The above image segmentation results are beneficial for extracting image-omics and image depth features and conducting multisource feature prediction in combination with clinical features.After feature selection and feature fusion are conducted,feature modeling and demonstration are performed based on machine learning and deep learning,respectively.Results and Promotion(1)Scale bridging single-dimensional segmentation diagnosis prediction: On internal data,2D-MAEN achieved 81.03% DSC and 82.19% F2,and 3D-MAEN achieved 78.75%DSC and 81.36% F2.On external data,2D-MAEN achieved 68.82% DSC and 65.36% F2,and 3D-MAEN achieved 70.75% DSC and 70.47% F2.DSC(%)values of 98.85,96.38,96.74,96.15,94.71,and 97.01 and F2(%)values of 98.64,96.80,98.01,97.47,94.00,and97.84 were obtained on the multisource data in order.DSC(%)values of 95.68,96.22,96.89,and 97.66 and F2(%)values of 94.85,96.34,96.25,and 97.77 were obtained for the multimode data.(2)Multidimensional segmentation diagnosis prediction of dimensional mixture: The partitioning performance of F-MHEN is superior to that of S-MHEN and T-MHEN.F-MHEN can attain 81.25% DSC and 83.88% F2 on internal data and 72.43% DSC and 76.67% F2 on external data.(3)Prognostic prediction of radiomics with feature fusion: In machine learning,the prediction model incorporating clinical and radiomics features can reach better prediction results,obtaining 92.54% AUC and 88.54% ACC.In deep learning,a deep ensemble learning model incorporating clinical and radiomic features can obtain 93.65% AUC and 93.75% ACC;a deep ensemble learning model incorporating multisource features(clinical,radiomics,and image depth)yielded 96.83% AUC and 95.56% ACC.ConclusionIn this paper,a "single-dimensional segmentation-multidimensional segmentationradiomics" deep learning system based on multisource data and various methods is created,and then the auxiliary diagnosis and prognosis evaluation of stroke are conducted.(1)In the auxiliary diagnosis,scale bridging single-dimensional segmentation and multidimensional segmentation with dimension blending are implemented in turn.In singledimensional segmentation,MAEN can effectively address the semantic gap problem.In multidimensional segmentation,F-MHEN can combine the advantages of 2D and 3D singledimensional segmentation to effectively enhance segmentation performance.(2)Multisource features of clinical,radiomics,and image depth were included in the prognostic assessment.The advantage of machine learning resides in its strong interpretability,which can further construct nomograms,decision curves,etc.The advantage of deep learning is that it has tremendous potential for predictive performance and can be combined with advanced techniques to achieve performance improvements.In summary,this paper can be condensed into the main line of application and promotion of "single-dimensional segmentation → multidimensional segmentation →radiomics" and successively employs the strategy of "scale bridging → dimension mixing →feature fusion".On the one hand,this paper analyzes the interconnection of image segmentation and radiomics.In addition,this paper provides an efficient and practicable solution for the diagnosis and prognosis of stroke.