Purpose: Most deep learning image registration networks predict the registration parameters only, and do not come with a “confidence” measure to alert a more intelligent human user for intervention. In this study, we attempt to address this issue by building a network that can associate confidence scores to its predictions.
Methods: The proposed network is based on the YOLO object detection pipeline which comes with a built-in confidence score. We modified the network to include a triplet loss function following the spirit of Siamese training that uses an intentionally chosen wrong image to improve feature embedding. The network was first trained over the PASCAL VOC dataset (22,471 images) to register a natural image to its blurred and shifted version. The resulting network was then finetuned over an in-house DRR-KV dataset (ten times smaller, 2,478 images) to register an online KV image to a spine-enhanced DRR template.
Results: After 900 epochs of training over the VOC dataset, the model reached a registration passing rate of 86.5% for a tolerance within 8 pixels (equivalent to 2mm for KV images). We observed that the model was able to consistently assign higher confidence scores to correct registrations which average to 0.76, as compared to 0.61 for the incorrect ones. For the in-house dataset, the above three metrics were 20.1%, 0.51, and 0.42, respectively. We again found consistent pattens that the algorithm assigns higher scores to correct predictions. Compare with VOC training, we believe a ten-fold increase of Xray dataset is necessary for the model to perform well over Xray images.
Conclusion: We developed a 2D rigid registration network that can quantify its own “certainty.” Preliminary results suggest it is a data-huger model. An improved version trained with more data has the potential to become a safer and more accurate AI platform for image registration.
Not Applicable / None Entered.