Automatic Craniomaxillofacial Landmark Digitization via Segmentation-Guided Partially-Joint Regression Forest Model and Multiscale Statistical Features

Abstract

Objective The goal of this paper is to automatically digitize craniomaxillofacial (CMF) landmarks efficiently and accurately from cone-beam computed tomography (CBCT) images, by addressing the challenge caused by large morphological variations across patients and image artifacts of CBCT images. Methods We propose a Segmentation-guided Partially-joint Regression Forest (S-PRF) model to automatically digitize CMF landmarks. In this model, a regression voting strategy is first adopted to localize each landmark by aggregating evidences from context locations, thus potentially relieving the problem caused by image artifacts near the landmark. Second, CBCT image segmentation is utilized to remove uninformative voxels caused by morphological variations across patients. Third, a partially-joint model is further proposed to separately localize landmarks based on the coherence of landmark positions to improve the digitization reliability. In addition, we propose a fast vector quantization (VQ) method to extract high-level multi-scale statistical features to describe a voxel's appearance, which has low dimensionality, high efficiency, and is also invariant to the local inhomogeneity caused by artifacts. Results Mean digitization errors for 15 landmarks, in comparison to the ground truth, are all less than 2mm. Conclusion Our model has addressed challenges of both inter-patient morphological variations and imaging artifacts. Experiments on a CBCT dataset show that our approach achieves clinically acceptable accuracy for landmark digitalization. Significance Our automatic landmark digitization method can be used clinically to reduce the labor cost and also improve digitalization consistency.

Detecting Anatomical Landmarks From Limited Medical Imaging Data Using Two-Stage Task-Oriented Deep Neural Networks

“…Given a 3D image patch, our goal is to learn a non-linear mapping to predict its 3D displacements to multiple landmarks. The conventional patch based landmark detection methods build the mapping using random forest regression models [14], [13], and usually require pre-defined appearance features to represent image patches. Without using any pre-defined features, we adopt a patch based regression model using CNN.…”

“…Due to variations of shape across subjects, the estimation of displacements for faraway landmarks from local image patches are often inaccurate [13]. Thus, we propose to adopt a weighted mean squared error as a loss function in the first-stage CNN model, by assigning lower weights to the displacements for faraway landmarks from image patches.…”

“…We first compare T 2 DL with two baseline methods, including 1) multi-atlas (MA) based method using non-linear registration [37], 2) random forest (RF) regression [38], [13]. We further compare our method with two state-of-the-art approaches, i.e.…”

“…In general, these methods can be roughly divided into three categories, including 1) keypoint based methods [11], 2) atlas based methods [12], and 3) learning based methods [13], [14], [15], [16], [17], [18]. Specifically, in keypoint based methods, points of interest (such as symmetry maxima, and Harris corners) detectors are adopted to identify candidates for individual model landmarks, and perform discrete optimization of a graph matching problem to obtain the final localization of landmarks.…”

“…Given plenty of image patches, the optimal landmark position can be estimated by assembling the predictions of a tremendous amount of image patches, e.g. , via a majority voting strategy [15], [13]. Therefore, context information from nearby patches can be utilized to localize landmarks.…”