Research On Classification Algorithms Of Medical Images For Tumors And Their Application In Hepatocellular Carcinoma

Medical image classification for tumors plays an important role in the diagnosis and treatment of medical diseases.With the development of physics,electronic engineering and computer science and technology,medical imaging has made great progress in imaging.Image acquisition speed is getting much faster,with higher image resolution and more image modality available.However,the interpretation of these images is still mainly dependent on imaging doctors.On the one hand,more images with higher quality greatly increase their burden;on the other hand,their interpretation of images relies mainly on qualitative image features that are visible to naked eyes,which is inevitably influenced by subjective factors such as personal experience.Therefore,how to obtain and utilize quantitative image features has always been an emphasis in relative research in order to achieve a more accurate classification of medical tumor images.This paper discusses the classification algorithms of medical tumor images and their applications in liver cancer,which are based on the physical nature and feature extraction methods of different medical images.The innovations of this paper are as follows:1)independently design a medical imaging 3D classification network called MviNet;2)apply the network to predict microvascular invasion(MVI)in patients with hepatocellular carcinoma(HCC)with accuracy,sensitivity,and specificity all above 0.7 and obtain twice the performance improvement in sensitivity;3)propose and develope multi-phase change rate features of hepatocellular carcinoma for the first time and achieve better results than previous reported studies when applied to predict MVI in patients with HCC.The specific research results are as follows:(1)A medical tumor image classification framework is proposed.It is mainly composed by integrating medical images(including image modality,basic physical properties of each modality and various imaging techniques corresponding to these modalities),quantitative image features and analytical methods.It emphasizes the importance of physical essence of medical imaging in medical image classification.Specifically,the apparent diffusion coefficient(ADC)map is selected as the image modality when discriminating cervical cancers from normal cervical tissues.Each first order statistic feature derived from ADC map can significantly differentiate them and the areas under the curve(AUC)of receiver operating characteristic for all those features exceed 0.85.By contract,when chosing the early,middle,late arterial phase and hepatobiliary phase of the Gd-EOB-DTPA dynamic enhanced magnetic resonance imaging to distinguish whether patients with HCC have MVI,each extracted whole-lesion first-order statistical and morphological feature in each phase is not sufficient to predict MVI,and AUC of all features do not reach 0.7.These two sets of tests indicate that for a specific medical classification problem,finding an appropriate kind of medical image reflecting the physical nature of the issue can greatly reduce the difficulty and complexity of classification,and verify the importance of physical nature of the issues and medical images concerned for medical tumor classification.(2)A classification method that integrates multi-phase and multi-feature of medical images is proposed.We extract the whole-lesion morphology and texture features of the early,middle,late arterial phase and hepatobiliary phase of the Gd-EOB-DTPA dynamic enhanced magnetic resonance image and develope the change rate feature group with the above four phases for the first time.Machine learning modals are built using the features extracted from each phase and the change rate feature group to predict MVI in patients with HCC.The results indicate that machine learning model based on the change rate feature group has the best performance.The AUC and the classification accuracy are both higher than 0.7 in test set,which is better than the previous reported studies.(3)A 3D classification network to predict MVI in patients with HCC is proposed.To further improve the classification efficiency,we carry out the following two studies in the early,middle,late arterial phase and hepatobiliary phase of the Gd-EOB-DTPA dynamic enhanced MRI,respectively,based on the data-driven approach:1)explore the transfer learning based on DenseNet-121,2)independently design a 3D classification network which is mainly built on the DenseNet and improve it in dimension,structure and type of features of the network based on medical image characteristics.In the dimension of the network,based on the fact that the tumor is a three-dimensional structure,the network is designed in three dimensions.In the structure of the network,we consider the following three aspects.Firstly,the input of the network is designed as a single channel since the medical image is usually single channel.Secondly,the direct connection between the feature maps of different layers of the network is established,resulting the more efficient utilization of the feature and avoiding the gradient vanishing.Thirdly,a 1 × 1 × 1 convolution layer is added before the fully connected layer to further reduce the number of parameters and avoid over-fitting.In the network feature type,the features learned by the convolutional network are often texture features.However,according to our previous research results,the morphological features are also helpful for distinguishing the microvascular invasion of HCC.Therefore,we add morphological features to the original network structure to enable an end-to-end classification network.Experiments show that the network trained on the patient’s hepatobiliary data achieves the best results,with test accuracy,sensitivity and specificity all higher than 70%,and in sensitivity it obtains twice the performance improvement,better than previous research results.