Unsupervised retinal image registration based on D-STUNet and progressive keypoint screening strategy.

Objective. Retinal image registration improves the accuracy and validity of a doctor's diagnosis and holds a crucial role in the monitoring and treatment of associated diseases. However, most existing image registration methods have limitations in identifying retinal vascular features, making it difficult to achieve desirable results in retinal image registration tasks. To solve this problem, a fusion network of Swin Transformer and U-Net, improved by Differential Multi-scale Convolutional Block Attention Module with Residual Mechanism (DMCR), named D-STUNet, is proposed in conjunction with the designed Progressive Keypoint Screening (PKS) strategy.

Approach: The D-STUNet network is primarily based on an encoder-decoder framework, and employs DMCR for the improvement and fusion of the Swin Transformer and U-Net networks. Among them, the DMCR module enhances the ability to focus on retinal vascular features, which effectively improves the accuracy of retinal image registration in the event of limited data. Simultaneously, the network introduces the PKS strategy to enable the gradual accumulation of effective keypoint information in the course of the training, which ensures that the keypoints are more concentrated in the retinal vascular region, thus enhancing the matching rate and overall detection effect.

Main results: The registration validation is conducted on the publicly accessible dataset Fundus Image Registration Dataset (FIRE) and compare it with nine algorithms. The experimental results show that the algorithm achieves an acceptance rate of 98.50%, a failure rate of 0, and an inaccuracy rate of 1.50%. In the area under the curve (AUC) metric, AUC for the Easy group is 0.929, while the AUC for the Mod and Hard groups are 0.883 and 0.724, respectively. The mean area under the curve (mAUC) across all comparison algorithms is the highest, outperforming the second-best algorithm by 0.09. Although it did not reach the optimum in certain subcategories (such as AUC-easy), its overall performance is significantly superior to existing methods.

Significance: The proposed network is able to effectively capture local features such as complex vascular structures in retinal images, providing a new method to improve the registration accuracy of retinal images.