Research On Fully Automatic Cardiac Magnetic Resonance Imaging Segmentation Based On Deep Learning

Cardiac magnetic resonance imaging(MRI)is mainly used for the evaluation of cardiac function parameters and the diagnosis of car-diovascular diseases,and is considered the gold standard for quanti-tative cardiac analysis.Obtaining accurate intracardiac structures from cardiac MRI is the first step in evaluating cardiac function pa-rameters.In clinical practice,doctors manually draw contour lines to calibrate intracardiac structures,but this method is time-consuming and has standard differences,which can lead to inaccurate segmenta-tion results.Therefore,a fully automatic heart segmentation algorithm is very necessary.With the rapid development of intelligent medical treatment,the real-time nature of the segmentation algorithm has also become very important,and the lightweight heart segmentation algo-rithm is also a current research hotspot.In this paper,after research and analysis of existing cardiac MRI image segmentation algorithms at home and abroad,two full-auto-matic cardiac MRI image segmentation algorithms based on deep learning are proposed.These are the heart MRI image segmentation algorithm based on densely connected holes and the lightweight heart MRI image segmentation algorithm based on separable convolution.The first algorithm combines dense connection and hollow convolu-tion to propose a hollow dense connection module.At the same time,a multi-scale fusion module and a channel attention mechanism mod-ule are introduced to improve model segmentation performance and extend the model to a three-dimensional segmentation algorithm.The 2D and 3D segmentation algorithms obtained the average Dice coef-ficient scores of 0.889 and 0.904 on the test set,respectively.The sec-ond algorithm uses YOLO-based ROI detection and U-Net-based im-proved segmentation network cooperation to reduce the number of model parameters and increase the model running speed.At the same time,it introduces residual connections to ensure the stability of the deep network,making segmentation Performance has been improved.The algorithm achieves an average Dice coefficient score of 0.904 on the CPU at an average speed of 357ms per image.This article conducts model training and testing based on short-axis movie sequence MRI images of 100 patients in the dataset of the Automated Cardiac Diagnostic Challenge(ACDC)held in 2017.Comparison with other algorithms and analysis of subjective image results verify the effectiveness of the proposed algorithm.