Left Ventricular Segmentation Framework Based on U-Net Localization and Improved Level Set Method

In medical image analysis application, some segmentation methods based on deep learning often make mistakes in segmentation on some fine edges, while the distance regularization level set method can achieve accurate segmentation in details, but the initial outline of segmentation needs to be determined artificially, and it is time-consuming and labor-intensive in clinical diagnosis and processing. The model proposed in this paper uses the U-Net network to preprocess the target image, and selects the inscribed rectangle of the processed result as the initial contour of the horizontal set, which solves the difficulty of initial contour selection. A new penalty function is proposed to avoid the curve from invading the target when the weak boundary is divided, and can maintain the property of the signed distance function well. The proposed new edge indication function and area term coefficient fuse the preprocessing results of U-Net network with the evolution curve, so that the evolution curve can adaptively adjust the convergence speed, improve the speed at the position of the pixel change moderately, and make the evolution curve accurately stay at the target boundary when approaching the target boundary, so as to avoid over-segmentation of the target image. Finally, the alternating direction multiplier method (ADMM) is used to solve the minimization of the energy function, which reduces the computational cost of the model and improves the computational efficiency. Experimental results on different target images show that the proposed model improves the accuracy of left ventricular segmentation, accelerates the segmentation speed, and has strong robustness.