Medical Ultrasound Image Segmentation Based on Improved MultiResUNet Network

An improved MultiResUNet is proposed to recognize tissue from clinical ultrasonic images. The designed anti-aliased convolution and MultiRes module is sensitive to high frequency semantics such as tissue boundaries, while the ResPath module reinforces the communication between encoding blocks and decoding blocks and reduces semantic discrepancies. During training process, a hybrid loss function is employed to focus on irregular or discontinuous boundaries. The result shows that proposed network outperforms conventional methods with similar parameter numbers and faster convergence speed.