Computer-aid Ventricle Segmentation For Cardiovascular Magnetic Resonance Images Based On Neural Network

With the continuous increase in the cases of cardiovascular disease in China and the development of medical imaging technology,medical imaging plays an important role in the clinical diagnosis and rehabilitation,but also brings considerable pressure to handle various medical datasets.Right ventricle(RV),as a part of heart,has noticeable difference between individuals.And the shape of RV may vary significantly under the different disease conditions and pathological stages.In addition,the technological difference among numerous equipment leads to diverse images.Those facts make it challenge to manually segment RV in cardiac images.Therefore,it is of vital importance to develop a robust and accurate computer-aided segmentation system to relieve the corresponding stress for experts or doctors.It will promote the clinical application of medical imaging,and further advance the early intervention of cardiovascular diseases.Based on the public cardiac magnetic resonance images,a research on RV segmentation was implemented as follows:ⅰ)In order to prevent over-fitting,we conducted data augmentation through the random combination of various image transformation preprocessing algorithms,including flipping,shifting,rotating and stretching.ⅱ)We considered full convolution neural network and typical U-Net network as the baseline model,to segment the dataset and discuss the problems in these models.ⅲ)A novel U-Net with dilated convolutional structure was constructed and trained on those datasets to explore the performance of dilated convectional structure,which contributes to expending the receptive field of deep neurons in a neural network.ⅳ)Residual and Dense networks with skip connections was introduced in U-shape network.We combined it with dilated convolutional structure to verify the robustness of such strategy.The results showed that the three components in Dense-U Net significantly contributed to ventricle segmentation task:ⅰ)The dilated convolutions in residual block were indeed effective for dealing with the holes appearing in the predicted masks and decrease false positive rate;ⅱ)The constant feature maps limited the exponential growth of parameters;ⅲ)The shortcuts in residual module and dense connection between input and each down-sampling block increased the performance and robustness of network.Finally,we accomplished a computer-aid system towards RV segmentation,in which the functions of data import,image preprocessing,neural model selection and evaluation as well as on-line segmentation were built.It provides one way for segment ventricles stably and quickly.