Research And Application Of Computer Aided Diagnosis Algorithm Based On Medical Image

With the advancement in the electronic science and technologies,more and more physical measurement methods have been developed.Specifically,in medical field,medical devices for multi-modality imaging have been put into practice.Medical images are of great value in clinic.Nowadays,imaging takes shorter time,possesses higher resolution and reflects more aspects of tissues.As a result,imaging becomes a routine clinical practice for detection and diagnosis.In contrast,it usually takes long time to educate radiologists,and the growth of the radiology department is slow.Long work hours and heavy workload inevitably result in missed diagnosis and misdiagnosis.Computer aided diagnosis(CAD)has been a hotspot in the interdisciplinary research of the computer science and medicine.It has been proved that it can provide effective decision support for doctors and improve the diagnostic accuracy under certain circumstances.Under this premise,the boosts of artificial intelligence(AI)makes it a new attempt that applying AI for more powerful medical imaging.Based on the clinical requirements,this study focuses on the algorithm and application of the CAD using medical images.There are three major innovations in this study.Firstly,based on the workflow of CAD,an in-house platform for CAD algorithm study using multi-modality images has been developed.Various innovative data pre-process and processing methods as well as statistics based and deep learning based algorithms have been integrated into the platform for CAD research.Secondly,this study explores the value of the whole-volume apparent diffusion coefficient(ADC)first-and second-order parameters in characterizing pathologic features in rectal cancer using statistical methods,showing the value of statistical methods in clinical feature classification as well as the potential of the ADC features in distinguishing rectal cancers from normal walls as well as in evaluating different pathologic factors including T stage and perineural invasion(PNI)of rectal cancers.Thirdly,in combination with the IBS theory,the cross convolutional neural network(CCNN)is proposed in this study to address the interpretability of the neural network as well as the issue of training on small datasets.In the aided diagnosis of liver diseases,the proposed CCNN improves the accuracy by at least 10%when compared with state-of-the-arts.Details are as follows.1.An in-house multi-modality medical image aided diagnosis platform is developed.Classical OSGi architecture is applied for the platform building,which make the platform stable and expandable.The platform well supports every step in the CAD workflow.Under the communication and feedback from the clinic,synchronized ROI contouring through multi-modality medical images is implemented in the platform,making ROI more accurate.More than 300 first-order,shape and second-order features can be extracted from images using the platform.Statistical methods as well as deep learning based algorithms are integrated in the platform.Statistics based methods are well defined in mathematics and can be interpretable,which is widely accepted by the clinic.On the other hand,deep learning based algorithms are data driven and learn from the data themselves instead of the manually-designed features,showing great potential in CAD.The two category of algorithms are complementary to make sure that the platform can be successfully applied in various scenarios.2.In the study of the rectal cancer pathologic features characterization using statistical methods,ADC features,including ADC Mean,10th,25th,50th,75th,90th percentile,Skewness,Kurtosis,Entropy and Entropy(H),derived from whole-lesion volumes of 50 patients are extracted using the above platform and compared between pathologic T1-2 and T3 stages,perineural invasion(PNI)present and absent,lymphangiovascular invasion(LVI)present and absent,and pathological NO and N+stages as well as between rectal cancers and normal walls.Statistical analyses including Wilcoxon test and Mann-Whitney U test are used to assess the differences between different groups.Receiver operating characteristic(ROC)analysis is conducted to evaluate the diagnostic performance and determine the optimal cut-off value of ADC parameters.Results show that all parameters indicate significant difference between malignant and benign rectal tissues,and Entropy and Entropy(H)boast largest areas under receiver operating characteristic(ROC)curve(AUC)of 0.98 and 0.97 respectively.Entropy and Entropy(H)are significantly lower in rectal cancers at T1-2 stages than T3 with AUCs of 0.78 and 0.83 respectively.90th percentile of rectal cancers with PNI is significantly lower than of those without PNI with an AUC of 0.74.All p<0.05.The findings demonstrate the potential of whole-lesion ADC features in characterizing rectal cancer pathologic factors using statistical algorithms.3.In combination with the IBS theory,cross convolutional neural network(CCNN)is proposed in this study.IBS is a special similarity or dissimilarity measurement used for measuring the difference between two feature maps to estimate the texture feature frequency distribution in the two original images,which measures the similarity in the original textures.We hypothesize that images of the same classification have similar texture features while images of different classifications have different texture features.In the CCNN,modified IBS is used as the loss function in the network.Once the CCNN is trained,the network can minimize the difference between images of the same classification and maximize the difference between images of different classfications.Compared with traditional NN,CCNN is well interpretable.CCNN extracts features from the paired images and measures their similarity based on IBS.Due to the special network architecture,in the prediction step,paired images consisting of one classification-known images and one or more images needing to be classified will be input into the network.As a result,the network can better take advantages of the prior knowledge learned from the training dataset without any specific classification-related hypothesis,making it feasible to train the network on a small dataset.VGG-19 network has been proved to be effective in feature extraction.The CCNN build its feature extraction part based on VGG-19 and is initialized using the pre-trained VGG-19 weights to save time for training and reduce the possibility of overfitting.4.The diagnosis performance of the above mentioned CCNN is explored in the study of the liver disease CAD and achieve good results,proving the excellent ability of the network to classify the lesions in liver.In details,the network can predict liver fibrosis stage with an accuracy of over 90%,predict pathologic factors of primary liver cancer with an accuracy of over 80%,indicating great improvement in diagnosis performance compared with previous studies.The findings show the great potential of deep learning based methods in medical image aided diagnosis.