Self-supervised local rotation-stable descriptors for 3D ultrasound registration using translation equivariant FCN

Rotation-stable descriptors are crucial for feature matching in medical image registration. Most existing descriptors rely on hand-crafted models to achieve rotation stability, which are susceptible to complex noise and fail to efficiently extract batches of three-dimensional features, particularly for ultrasound volume. In this study, a translation equivariant design was performed based on the fully convolutional network to extract descriptors at different positions in batches by removing position bias errors, thereby improving the descriptor extraction efficiency. Descriptor rotation consistency is used for self-supervised training to avoid the need for data annotation. Before matching, the image ROI is restructured to adjust the input size of the network, further improving the descriptor extraction efficiency. Then, the multi-consistencies filter based on the correlation among descriptors, spatial positions, and texture features is designed to preserve stable matched pairs for accurate and robust registration results. Classification experimental results based on rotation stability show that the descriptors extracted by the proposed method have high classification accuracy, particularly under interference, such as noise, blur, and artifacts. Experimental results of clinical ultrasound image registration show that the proposed method has a lower registration error of 3.59 ± 1.15 mm compared with other methods. In addition, the descriptor extraction network proposed in this study has low training costs and high processing speed, further revealing the potential of the proposed method in clinical applications.