Research on Heart Image Segmentation Algorithm Based on Improved UNet++Model

Cardiac segmentation is of great significance in medical image analysis, providing critical support for the diagnosis, treatment, and postoperative evaluation of cardiovascular diseases. However, due to the complex morphology of the heart and differences in patient anatomy, high-precision heart segmentation faces challenges. This article proposes an improved UNet++ model. Firstly, by replacing the ordinary convolutional layer in the UNet++ model with DenseASPP, the receiving field is extended during the downsampling stage, effectively avoiding the problem of feature loss caused by increasing network depth. Secondly, by utilizing ECA channel attention to automatically select important information extracted from heart features by deep networks, good segmentation performance can be achieved. Finally, train the improved UNet++ model to achieve end-to-end heart image segmentation. The experiment used the ACDC cardiac dataset, with an evaluation index F1 score and an average Dice of 91.79%, and an average recall rate of 91.74%.