Dynamic U-shaped convolutional network for mouse cardiac image segmentation and quantification.

Accurately segmenting and quantifying mouse cardiac slice images with myocardial infarction is of great significance in cardiovascular disease research. Manual methods are time-consuming, so automatic segmentation is highly sought after. However, due to the irregular U-shaped structure of infarcted and high-risk areas, the task is complex. This study proposes a dynamic U-shaped convolutional network that adapts to these irregular structures. We designed a dynamic convolution to focus on U-shaped local features, and developed a dual-stream fusion block to enhance performance. An attention gate mechanism suppresses irrelevant information, highlighting key features. We constructed a dataset of 243 mouse cardiac slice images with myocardial infarction. The proposed method outperforms other models, achieving an average Dice coefficient of 80.68%, 1.7% higher than the best existing algorithm. For infarct size segmentation, it reached 80.13%, surpassing the current optimal method by 2.43%. Additionally, our approach can quantify the ratio of infarcted to risk areas, aiding the assessment of myocardial injury severity. The dataset is available at https://github.com/wwz4416/mouse-cardiac-dataset.