LandmarkNet: a 2D digital radiograph landmark estimator for registration.

BACKGROUND: Radiation therapy requires precision to target and escalate the doses to affected regions while reducing the adjacent normal tissue exposed to high radiotherapy doses. Image guidance has become the start of the art in the treating process. Registering the digital radiographs megavoltage x ray (MV-DRs) and the kilovoltage digital reconstructed radiographs (KV-DRRs) is difficult because of the poor quality of MV-DRs. We simplify the problem by registering between landmarks instead of entire image information, thence we propose a model to estimate the landmark accurately. METHODS: After doctors' analysis, it is proved that it is effective to register through several physiological features such as spinous process, tracheal bifurcation, Louis angle. We propose the LandmarkNet, a novel keypoint estimation architecture, can automatically detect keypoints in blurred medical images. The method applies the idea of Feature Pyramid Network (FPN) twice to merge the cross-scale and cross-layer features for feature extraction and landmark estimation successively. Intermediate supervision is used at the end of the first FPN to ensure that the underlying parameters are updated normally. The network finally produces heatmap to display the approximate location of landmarks and we obtain accurate position estimation after non-maximum suppression (NMS) processing. RESULTS: Our method could obtain accurate landmark estimation in the dataset provided by several cancer hospitals and labeled by ourselves. The standard percentage of correct keypoints (PCK) within 8 pixels of estimation for the spinous process, tracheal bifurcation and Louis angle is 81.24%, 98.95% and 85.61% respectively. For the above three landmarks, the mean deviation between the predicted location of each landmark and corresponding ground truth is 2.38, 0.98 and 2.64 pixels respectively. CONCLUSION: Landmark estimation based on LandmarkNet has high accuracy for different kinds of landmarks. Our model estimates the location of tracheal bifurcation especially accurately because of its obvious features. For the spinous process, our model performs well in quantity estimation as well as in position estimation. The wide application of our method assists doctors in image-guided radiotherapy (IGRT) and provides the possibility of precise treatment in the true sense.