Research On Medical Image Classification Using Deep Features In CT And X-Rays Images

Medical image classification plays a great role in clinical treatment.Deep learning techniques,especially convolutional neural networks(CNNs),have proven potential for several classification tasks and visual object recognition.Hоwever,in the field of medical image classification,соnventiоnаl deeр leаrning techniques hаve twо mаin drаwbасks: 1)insuffiсient trаining dаtа аnd 2)the dоmаin mismаtсh between the trаining dаtа аnd the testing dаtа.We approach these challenges with a presentation of three novel methods.The principal works are as follows:The first approach realizes a classification of four types of liver lesions,namely,hepatocellular carcinoma,metastases,hemangiomas,and healthy tissues based on convolutional neural networks with a concise model named Fire Net.The proposed Fire Net introduced fire modules from SqueezeNet to reduce the number of parameters and the model size while enhancing speed for quick classification.A convolution layer of 1x1 is added on top of each concatenation for preserving the feature information of different layers used for classification.Bypass connection was added around fire modules for learning a residual function between input and output,and to solve the vanishing gradient problem.Once our proposed model has a small number of parameters,training on small datasets becomes feasible.In addition,a new particle swarm optimization(NPSO)is proposed to optimize the network parameters in order to further improve the performance of the proposed Fire Net model.The performance of the proposed model is analyzed and compared with the existing methods.From the analysis,it is observed that the proposed Fire Net with NPSO achieved 89.2% classification accuracy,which is higher classification accuracy than other methods.The second approach is an extension to the first approach.It combines two novel neural networks for classifying liver lesions of Hepatocellular carcinoma,Metastases,Hemangiomas,and Healthy tissues.We used Fire Net as the first model.In this part,Fire Net introduces fire modules to reduce the model size and the number of parameters for quick classification.Fire Net will be in charge of extracting feature information.The second model named Modified Long Short-Term Memory(MLstm).In this part,the features extracted by the Fire Net are then used for temporal information and prediction with a new loss function named G-loss.In order to improve the proposed Fire Net-MLstm model,a new bias was added through the forget gate.The hyperbolic tangent and the activation function were also added.The proposed model has achieved an accuracy of up to 91.2% in classifying liver lesions.Finally,we proposed a novel distant domain transfer learning(DDTL),called feature fusion,decompose and transfer(FFDT)for classifying COVID-19,normal,and pneumonia using X-ray images.We leverage on transfer learning from distant source domain to obtain more information for the target task.Moreover,the proposed FFDT method has three parts: 1)The proposed FFDT benefits from fusing distant features extracted from distant domains into a common feature space where the distribution mismatch is minimized.2)We used Fire Net method to extract feature using the class reconstruction to unravel the local structure of the data distribution.3)We explored the effect of adding the fully connected layer after knowledge transformation to build a target classifier for target task.Extensive exрeriments аreрerfоrmed tо evаluаte the рerfоrmаnсe оf the рrороsed FFDT methоd оn СОVID-19 X-rаy imаges fоr tаrget dоmаin аnd Саlteсh-256 fоr sоurсe dоmаin.As a result,the proposed FFDT method brings about 94.5% classification accuracy,which is superior to the other comparative methods.In summary,the proposed Fire Net based on the new particle swarm optimization algorithm,the modified long short term memory(MLstm)and a new distant domain transfer learning(DDTL)method proposed in this dissertation have a good effect on medical image disease classification and achieve the expected research goal.