For dual source CT image with contrast media, due to heart soft tissue density andcontrast media uneven distribution result in the CT value of heart tissues uneven and boundary fuzzy, taking a single image segmentation algorithm is too difficult to obtain satisfactory results, so morphological reconstruction and random walks hybrid method is proposed in this paper.Firstly，we used morphological reconstruction operation on image smoothing filtering,which makes the heart cavity gray information convergence and gray level differences with the surrounding tissue and get the left atrium area with the fuzzy boundary;Then the random walks algorithm sets the seed points for each region of the image, and gives the weight of each side, and takes the weight of the edge as the transfer probability. For each unlabeled point is calculated from the point of first arrival probability of seed points.Finally, according to the first hit probability to choose the maximum that a class as belonging to the class, attribute of the unlabeled points and finally get the accurate left atrial.