Research And Application Of Medical Image Analysis And Cognitive Computing Methods Based On Deep Learning

Medical image computer-aided diagnosis has always been a research focus in computer application.The controllability of medical image acquisition and the need of clinical diagnosis provide application scenarios for deep learning.This paper mainly discusses a deep learning solution based on multi-parametric MRI images for non-invasive assessment of hepatocellular carcinoma(HCC)differentiation.Multi-channel three-dimensional convolutional neural networks and multi-scale deep residual networks are proposed to extract features of three-dimensional medical image data and two-dimensional fusion data.At the same time,the role of transfer learning and metric learning in the classification of medical images was verified.The main contributions of this paper are as follows:(1)According to the research results of domestic and foreign scholars and the doctor's clinical diagnosis experience,Multi-parametric MRI are collected and labeled as the materials of our experiments.The medical image dataset used to train the machine learning model includes T2 weighted imaging(T2WI),chemical shift imaging(inphase and out of phase)and dynamic contrast-enhanced magnetic resonance imaging(DCE-MRI).Data enhancement and sample resampling methods were used to improve the sample size and sample imbalance.At the same time,according to the collected medical images,the time intensity curve with important clinical diagnostic value is counted,which provides an experimental basis for establishing an effective examination plan and preferred image acquisition method.(2)A multi-channel fusion three-dimensional convolutional neural network architecture is proposed to extract the temporal-spatial features of DCE-MRI images.According to the data characteristics of spatial information and time series information of DCE-MR images,a tensor-based data representation model and a data fusion model are proposed to explore the effects of different data representation methods and network structures in image recognition computing.According to the results of the comparative experiments,the extraction of DCE-MRI time series information has better classification results in HCC and cirrhosis background differentiation and HCC differentiation degree assessment tasks.In the aspect of data fusion strategy,a multi-channel input fusion 3DCNN and feature splicing fusion network are proposed.The data is merged at the data level and feature level to extract temporal and spatial characteristics of DCEMRI images.The accuracy of medical image classification increased by 7.3%.(3)A deep learning model of medical image-assisted diagnosis was established by using data fusion,migration learning and multi-scale feature extraction.According to the complementarity of multi-modality images in diagnosis decision-making,when using multiple modal data,finding complementary modal data for fusion can effectively improve the diagnostic effect;Experiments have confirmed that although there are obvious differences between natural images and medical images,using the pre-trained model on the natural image dataset as the initialization of the network can ensure and accelerate the convergence of the training,and improve the performance on the testing set.(4)With the idea of metric learning,the classification problem of medical images is transformed into the measurement problem of similarity between medical image samples,and the effectiveness of the few-shot learning method in medical image classification is verified.The metric learning method guides the feature extraction process of the medical image,and reduces the distance of the individual data in the feature space.