Right Ventricular Segmentation Of Magnetic Resonance Imaging Based On Deep Leaning

In recent years,with the fast development of medical instruments,the method that doctors judge patients’ health conditions through medical images has been widely used.The results of the segmentation of target organs and tissues in medical images can effectively help doctors know patients’ state of illness.However,facing of massive medical images,if analyzing medical images through manual methods,doctors can’t satisfy the more and more requests of patients in speed and accuracy.So,effective and precise segnentation method of medical images is vital in Clinical medicine.With the successful application of deep learning in the field of speech signal processing and image signal processing,this technology has been transplanted into the field of medical image segmentation,and its segmentation result and speed are superior to traditional algorithms.In this paper,I investigate various existing deep neural network models,integrate advantages of existing technologies,and propose a new model called DCFD一Net(Dense connection of fully dilated-convolution network)for right ventricular segmentation in cardiac MR images(Cardiac magnetic resonance images).In recent years,the mainstream method of neural network in the field of image segmentation is to extract high level features through down-sampling and restore images through up-sampling.This method will result in reduced image segmentation accuracy and poor edge segmentation.I propose a new method that fusing extracting feature path and restoring image path to make up for the shortcomings of traditional neural network segmentation methods in this paper.But my method still have defects like rough edge segmentation and odd segmentation points.So in this paper,I make corresponding improvements in the aspects of training set,loss function,model optimizer and so on.I also add a post processing method at the end of my model to improve the segmentation result.Taking the famous U-Net network as the baseline model,this paper compares the proposed DCFD-Net with this.Compared with U-Net,DCFD-Net has enhanced edge segmentation ability and small target segmentation ability,and has improved on both the Dice similarity coefficient and the Jaccard similarity coefficient.It is proved that the DCFD-Net proposed in this paper has better result.DCFD-Net achieved 89.97%Dice coefficient and 82.72%Jaccard coefficient on the test set,which is slightly different from the artificially reproduced 93%-95%Dice coefficient,but the segmentation speed is far Faster than manual reproduction,and it has promoted the clinical application of neural networks.