Multimodality image registration by maximization of mutual information

A new approach to the problem of multimodality medical image registration is proposed, using a basic concept from information theory, mutual information (MI), or relative entropy, as a new matching criterion. The method presented in this paper applies MI to measure the statistical dependence or information redundancy between the image intensities of corresponding voxels in both images, which is assumed to be maximal if the images are geometrically aligned. Maximization of MI is a very general and powerful criterion, because no assumptions are made regarding the nature of this dependence and no limiting constraints are imposed on the image content of the modalities involved. The accuracy of the MI criterion is validated for rigid body registration of computed tomography (CT), magnetic resonance (MR), and photon emission tomography (PET) images by comparison with the stereotactic registration solution, while robustness is evaluated with respect to implementation issues, such as interpolation and optimization, and image content, including partial overlap and image degradation. Our results demonstrate that subvoxel accuracy with respect to the stereotactic reference solution can be achieved completely automatically and without any prior segmentation, feature extraction, or other preprocessing steps which makes this method very well suited for clinical applications.

Automated model-based tissue classification of MR images of the brain

We describe a fully automated method for model-based tissue classification of magnetic resonance (MR) images of the brain. The method interleaves classification with estimation of the model parameters, improving the classification at each iteration. The algorithm is able to segment single- and multispectral MR images, corrects for MR signal inhomogeneities, and incorporates contextual information by means of Markov random Fields (MRF's). A digital brain atlas containing prior expectations about the spatial location of tissue classes is used to initialize the algorithm. This makes the method fully automated and therefore it provides objective and reproducible segmentations. We have validated the technique on simulated as well as on real MR images of the brain.

Comparison and evaluation of retrospective intermodality brain image registration techniques

PURPOSE

The primary objective of this study is to perform a blinded evaluation of a group of retrospective image registration techniques using as a gold standard a prospective, marker-based registration method. To ensure blindedness, all retrospective registrations were performed by participants who had no knowledge of the gold standard results until after their results had been submitted. A secondary goal of the project is to evaluate the importance of correcting geometrical distortion in MR images by comparing the retrospective registration error in the rectified images, i.e., those that have had the distortion correction applied, with that of the same images before rectification.

METHOD

Image volumes of three modalities (CT, MR, and PET) were obtained from patients undergoing neurosurgery at Vanderbilt University Medical Center on whom bone-implanted fiducial markers were mounted. These volumes had all traces of the markers removed and were provided via the Internet to project collaborators outside Vanderbilt, who then performed retrospective registrations on the volumes, calculating transformations from CT to MR and/ or from PET to MR. These investigators communicated their transformations again via the Internet to Vanderbilt, where the accuracy of each registration was evaluated. In this evaluation, the accuracy is measured at multiple volumes of interest (VOIs), i.e., areas in the brain that would commonly be areas of neurological interest. A VOI is defined in the MR image and its centroid c is determined. Then, the prospective registration is used to obtain the corresponding point c' in CT or PET. To this point, the retrospective registration is then applied, producing c" in MR. Statistics are gathered on the target registration error (TRE), which is the distance between the original point c and its corresponding point c".

RESULTS

This article presents statistics on the TRE calculated for each registration technique in this study and provides a brief description of each technique and an estimate of both preparation and execution time needed to perform the registration.

CONCLUSION

Our results indicate that retrospective techniques have the potential to produce satisfactory results much of the time, but that visual inspection is necessary to guard against large errors.

Automated model-based bias field correction of MR images of the brain

We propose a model-based method for fully automated bias field correction of MR brain images. The MR signal is modeled as a realization of a random process with a parametric probability distribution that is corrupted by a smooth polynomial inhomogeneity or bias field. The method we propose applies an iterative expectation-maximization (EM) strategy that interleaves pixel classification with estimation of class distribution and bias field parameters, improving the likelihood of the model parameters at each iteration. The algorithm, which can handle multichannel data and slice-by-slice constant intensity offsets, is initialized with information from a digital brain atlas about the a priori expected location of tissue classes. This allows full automation of the method without need for user interaction, yielding more objective and reproducible results. We have validated the bias correction algorithm on simulated data and we illustrate its performance on various MR images with important field inhomogeneities. We also relate the proposed algorithm to other bias correction algorithms.

Automated segmentation of multiple sclerosis lesions by model outlier detection

This paper presents a fully automated algorithm for segmentation of multiple sclerosis (MS) lesions from multispectral magnetic resonance (MR) images. The method performs intensity-based tissue classification using a stochastic model for normal brain images and simultaneously detects MS lesions as outliers that are not well explained by the model. It corrects for MR field inhomogeneities, estimates tissue-specific intensity models from the data itself, and incorporates contextual information in the classification using a Markov random field. The results of the automated method are compared with lesion delineations by human experts, showing a high total lesion load correlation. When the degree of spatial correspondence between segmentations is taken into account, considerable disagreement is found, both between expert segmentations, and between expert and automatic measurements.

Modeling 3D facial shape from DNA

Human facial diversity is substantial, complex, and largely scientifically unexplained. We used spatially dense quasi-landmarks to measure face shape in population samples with mixed West African and European ancestry from three locations (United States, Brazil, and Cape Verde). Using bootstrapped response-based imputation modeling (BRIM), we uncover the relationships between facial variation and the effects of sex, genomic ancestry, and a subset of craniofacial candidate genes. The facial effects of these variables are summarized as response-based imputed predictor (RIP) variables, which are validated using self-reported sex, genomic ancestry, and observer-based facial ratings (femininity and proportional ancestry) and judgments (sex and population group). By jointly modeling sex, genomic ancestry, and genotype, the independent effects of particular alleles on facial features can be uncovered. Results on a set of 20 genes showing significant effects on facial features provide support for this approach as a novel means to identify genes affecting normal-range facial features and for approximating the appearance of a face from genetic markers.

Comparative evaluation of multiresolution optimization strategies for multimodality image registration by maximization of mutual information

Maximization of mutual information of voxel intensities has been demonstrated to be a very powerful criterion for three-dimensional medical image registration, allowing robust and accurate fully automated affine registration of multimodal images in a variety of applications, without the need for segmentation or other preprocessing of the images. In this paper, we investigate the performance of various optimization methods and multiresolution strategies for maximization of mutual information, aiming at increasing registration speed when matching large high-resolution images. We show that mutual information is a continuous function of the affine registration parameters when appropriate interpolation is used and we derive analytic expressions of its derivatives that allow numerically exact evaluation of its gradient. Various multiresolution gradient- and non-gradient-based optimization strategies, such as Powell, simplex, steepest-descent, conjugate-gradient, quasi-Newton and Levenberg-Marquardt methods, are evaluated for registration of computed tomography (CT) and magnetic resonance images of the brain. Speed-ups of a factor of 3 on average compared to Powell's method at full resolution are achieved with similar precision and without a loss of robustness with the simplex, conjugate-gradient and Levenberg-Marquardt method using a two-level multiresolution scheme. Large data sets such as 256(2) x 128 MR and 512(2) x 48 CT images can be registered with subvoxel precision in <5 min CPU time on current workstations.

Genome-wide mapping of global-to-local genetic effects on human facial shape

Genome-wide association scans of complex multipartite traits like the human face typically use preselected phenotypic measures. Here we report a data-driven approach to phenotyping facial shape at multiple levels of organization, allowing for an open-ended description of facial variation while preserving statistical power. In a sample of 2,329 persons of European ancestry, we identified 38 loci, 15 of which replicated in an independent European sample (n = 1,719). Four loci were completely new. For the others, additional support (n = 9) or pleiotropic effects (n = 2) were found in the literature, but the results reported here were further refined. All 15 replicated loci highlighted distinctive patterns of global-to-local genetic effects on facial shape and showed enrichment for active chromatin elements in human cranial neural crest cells, suggesting an early developmental origin of the facial variation captured. These results have implications for studies of facial genetics and other complex morphological traits.

Large-scale in-vivo Caucasian facial soft tissue thickness database for craniofacial reconstruction

A large-scale study of facial soft tissue depths of Caucasian adults was conducted. Over a 2-years period, 967 Caucasian subjects of both sexes, varying age and varying body mass index (BMI) were studied. A user-friendly and mobile ultrasound-based system was used to measure, in about 20min per subject, the soft tissue thickness at 52 facial landmarks including most of the landmarks used in previous studies. This system was previously validated on repeatability and accuracy [S. De Greef, P. Claes, W. Mollemans, M. Loubele, D. Vandermeulen, P. Suetens, G. Willems, Semi-automated ultrasound facial soft tissue depth registration: method and validation. J. Forensic Sci. 50 (2005)]. The data of 510 women and 457 men were analyzed in order to update facial soft tissue depth charts of the contemporary Caucasian adult. Tables with the average thickness values for each landmark as well as the standard deviation and range, tabulated according to gender, age and BMI are reported. In addition, for each landmark and for both sexes separately, a multiple linear regression of thickness versus age and BMI is calculated. The lateral asymmetry of the face was analysed on an initial subset of 588 subjects showing negligible differences and thus warranting the unilateral measurements of the remaining subjects. The new dataset was statistically compared to three datasets for the Caucasian adults: the traditional datasets of Rhine and Moore [J.S. Rhine, C.E. Moore, Tables of facial tissue thickness of American Caucasoids in forensic anthropology. Maxwell Museum Technical series 1 (1984)] and Helmer [R. Helmer, Schädelidentifizierung durch elektronische bildmischung, Kriminalistik Verlag GmbH, Heidelberg, 1984] together with the most recent in vivo study by Manhein et al. [M.H. Manhein, G.A. Listi, R.E. Barsley, R. Musselman, N.E. Barrow, D.H. Ubelbaker, In vivo facial tissue depth measurements for children and adults. J. Forensic Sci. 45 (2000) 48-60]. The large-scale database presented in this paper offers a denser sampling of the facial soft tissue depths of a more representative subset of the actual Caucasian population over the different age and body posture subcategories. This database can be used as an updated chart for manual and computer-based craniofacial approximation and allows more refined analyses of the possible factors affecting facial soft tissue depth.

Optimization for Medical Image Segmentation: Theory and Practice When Evaluating With Dice Score or Jaccard Index

In many medical imaging and classical computer vision tasks, the Dice score and Jaccard index are used to evaluate the segmentation performance. Despite the existence and great empirical success of metric-sensitive losses, i.e. relaxations of these metrics such as soft Dice, soft Jaccard and Lovász-Softmax, many researchers still use per-pixel losses, such as (weighted) cross-entropy to train CNNs for segmentation. Therefore, the target metric is in many cases not directly optimized. We investigate from a theoretical perspective, the relation within the group of metric-sensitive loss functions and question the existence of an optimal weighting scheme for weighted cross-entropy to optimize the Dice score and Jaccard index at test time. We find that the Dice score and Jaccard index approximate each other relatively and absolutely, but we find no such approximation for a weighted Hamming similarity. For the Tversky loss, the approximation gets monotonically worse when deviating from the trivial weight setting where soft Tversky equals soft Dice. We verify these results empirically in an extensive validation on six medical segmentation tasks and can confirm that metric-sensitive losses are superior to cross-entropy based loss functions in case of evaluation with Dice Score or Jaccard Index. This further holds in a multi-class setting, and across different object sizes and foreground/background ratios. These results encourage a wider adoption of metric-sensitive loss functions for medical segmentation tasks where the performance measure of interest is the Dice score or Jaccard index.

A unifying framework for partial volume segmentation of brain MR images

Accurate brain tissue segmentation by intensity-based voxel classification of magnetic resonance (MR) images is complicated by partial volume (PV) voxels that contain a mixture of two or more tissue types. In this paper, we present a statistical framework for PV segmentation that encompasses and extends existing techniques. We start from a commonly used parametric statistical image model in which each voxel belongs to one single tissue type, and introduce an additional downsampling step that causes partial voluming along the borders between tissues. An expectation-maximization approach is used to simultaneously estimate the parameters of the resulting model and perform a PV classification. We present results on well-chosen simulated images and on real MR images of the brain, and demonstrate that the use of appropriate spatial prior knowledge not only improves the classifications, but is often indispensable for robust parameter estimation as well. We conclude that general robust PV segmentation of MR brain images requires statistical models that describe the spatial distribution of brain tissues more accurately than currently available models.

Automatic 3-D segmentation of internal structures of the head in MR images using a combination of similarity and free-form transformations: Part I, Methodology and validation on normal subjects

The study presented in this paper tests the hypothesis that the combination of a global similarity transformation and local free-form deformations can be used for the accurate segmentation of internal structures in MR images of the brain. To quantitatively evaluate our approach, the entire brain, the cerebellum, and the head of the caudate have been segmented manually by two raters on one of the volumes (the reference volume) and mapped back onto all the other volumes, using the computed transformations. The contours so obtained have been compared to contours drawn manually around the structures of interest in each individual brain. Manual delineation was performed twice by the same two raters to test inter- and intrarater variability. For the brain and the cerebellum, results indicate that for each rater, contours obtained manually and contours obtained automatically by deforming his own atlas are virtually indistinguishable. Furthermore, contours obtained manually by one rater and contours obtained automatically by deforming this rater's own atlas are more similar than contours obtained manually by two raters. For the caudate, manual intra- and interrater similarity indexes remain slightly better than manual versus automatic indexes, mainly because of the spatial resolution of the images used in this study. Qualitative results also suggest that this method can be used for the segmentation of more complex structures, such as the hippocampus.

A viscous fluid model for multimodal non-rigid image registration using mutual information

We propose a multimodal free-form registration algorithm based on maximization of mutual information. The warped image is modeled as a viscous fluid that deforms under the influence of forces derived from the gradient of the mutual information registration criterion. Parzen windowing is used to estimate the joint intensity probability of the images to be matched. The method is evaluated for non-rigid inter-subject registration of MR brain images. The accuracy of the method is verified using simulated multi-modal MR images with known ground truth deformation. The results show that the root mean square difference between the recovered and the ground truth deformation is smaller than 1 voxel. We illustrate the application of the method for atlas-based brain tissue segmentation in MR images in case of gross morphological differences between atlas and patient images.

Nonrigid image registration using conditional mutual information

Maximization of mutual information (MMI) is a popular similarity measure for medical image registration. Although its accuracy and robustness has been demonstrated for rigid body image registration, extending MMI to nonrigid image registration is not trivial and an active field of research. We propose conditional mutual information (cMI) as a new similarity measure for nonrigid image registration. cMI starts from a 3-D joint histogram incorporating, besides the intensity dimensions, also a spatial dimension expressing the location of the joint intensity pair. cMI is calculated as the expected value of the cMI between the image intensities given the spatial distribution. The cMI measure was incorporated in a tensor-product B-spline nonrigid registration method, using either a Parzen window or generalized partial volume kernel for histogram construction. cMI was compared to the classical global mutual information (gMI) approach in theoretical, phantom, and clinical settings. We show that cMI significantly outperforms gMI for all applications.

Automatic analysis of cerebral asymmetry: an exploratory study of the relationship between brain torque and planum temporale asymmetry

Leftward occipital and rightward frontal lobe asymmetry (brain torque) and leftward planum temporale asymmetry have been consistently reported in postmortem and in vivo neuroimaging studies of the human brain. Here automatic image analysis techniques are applied to quantify global and local asymmetries, and investigate the relationship between brain torque and planum temporale asymmetries on T1-weighted magnetic resonance (MR) images of 30 right-handed young healthy subjects (15 male, 15 female). Previously described automatic cerebral hemisphere extraction and 3D interhemispheric reflection-based methods for studying brain asymmetry are applied with a new technique, LowD (Low Dimension), which enables automatic quantification of brain torque. LowD integrates extracted left and right cerebral hemispheres in columns orthogonal to the midsagittal plane (2D column maps), and subsequently integrates slices along the brain's anterior-posterior axis (1D slice profiles). A torque index defined as the magnitude of occipital and frontal lobe asymmetry is computed allowing exploratory investigation of relationships between this global asymmetry and local asymmetries found in the planum temporale. LowD detected significant torque in the 30 subjects with occipital and frontal components found to be highly correlated (P<0.02). Significant leftward planum temporale asymmetry was detected (P<0.05), and the torque index correlated with planum temporale asymmetry (P<0.001). However, torque and total brain volume were not correlated. Therefore, although components of cerebral asymmetry may be related, their magnitude is not influenced by total hemisphere volume. LowD provides increased sensitivity for detection and quantification of brain torque on an individual subject basis, and future studies will apply these techniques to investigate the relationship between cerebral asymmetry and functional laterality.

Spatially-dense 3D facial asymmetry assessment in both typical and disordered growth

Mild facial asymmetries are common in typical growth patterns. Therefore, detection of disordered facial growth patterns in individuals characterized by asymmetries is preferably accomplished by reference to the typical variation found in the general population rather than to some ideal of perfect symmetry, which rarely exists. This presents a challenge in developing an asymmetry assessment tool that is applicable, without modification, to detect both mild and severe facial asymmetries. In this paper we use concepts from geometric morphometrics to obtain robust and spatially-dense asymmetry assessments using a superimposition protocol for comparison of a face with its mirror image. Spatially-dense localization of asymmetries was achieved using an anthropometric mask consisting of uniformly sampled quasi-landmarks that were automatically indicated on 3D facial images. Robustness, in the sense of an unbiased analysis under increasing asymmetry, was ensured by an adaptive, robust, least-squares superimposition. The degree of overall asymmetry in an individual was scored using a root-mean-squared-error, and the proportion was scored using a novel relative significant asymmetry percentage. This protocol was applied to a database of 3D facial images from 359 young healthy individuals and three individuals with disordered facial growth. Typical asymmetry statistics were derived and were mainly located on, but not limited to, the lower two-thirds of the face in males and females. The asymmetry in males was more extensive and of a greater magnitude than in females. This protocol and proposed scoring of asymmetry with accompanying reference statistics will be useful for the detection and quantification of facial asymmetry in future studies.

On the problem of geometric distortion in magnetic resonance images for stereotactic neurosurgery

In this paper, we discuss the issue of geometric distortion in magnetic resonance (MR) images used to plan stereotactic neurosurgical interventions. We analyze the process for the case of Fourier transform imaging and demonstrate that spatial misregistrations are fundamentally due to two causes: deviations of the magnetic field from its ideal value and blood flow. This enables us to relate the causes of geometric distortion to the MR imaging system, the patient and the stereotactic localizer frame. Based on the general model, we propose model refinements and discuss methods for the quantification and correction of all causes. The results of our calculations and experiments indicate that, using the proposed corrections, MRI and MR angiography should be considered valuable and reliable acquisition modalities for the planning of stereotactic neurosurgical interventions.

Craniofacial reconstruction using a combined statistical model of face shape and soft tissue depths: methodology and validation

Forensic facial reconstruction aims at estimating the facial outlook associated with an unidentified skull specimen. Estimation is generally based on tabulated average values of soft tissue thicknesses measured at a sparse set of landmarks on the skull. Traditional 'plastic' methods apply modeling clay or plasticine on a cast of the skull, approximating the estimated tissue depths at the landmarks and interpolating in between. Current computerized techniques mimic this landmark interpolation procedure using a single static facial surface template. However, the resulting reconstruction is biased by the specific choice of the template and no face-specific regularization is used during the interpolation process. We reduce the template bias by using a flexible statistical model of a dense set of facial surface points, combined with an associated sparse set of skull-based landmarks. This statistical model is constructed from a facial database of (N = 118) individuals and limits the reconstructions to statistically plausible outlooks. The actual reconstruction is obtained by fitting the skull-based landmarks of the template model to the corresponding landmarks indicated on a digital copy of the skull to be reconstructed. The fitting process changes the face-specific statistical model parameters in a regularized way and interpolates the remaining landmark fit error using a minimal bending thin-plate spline (TPS)-based deformation. Furthermore, estimated properties of the skull specimen (BMI, age and gender, e.g.) can be incorporated as conditions on the reconstruction by removing property-related shape variation from the statistical model description before the fitting process. The proposed statistical method is validated, both in terms of accuracy and identification success rate, based on leave-one-out cross-validation tests applied on the facial database. Accuracy results are obtained by statistically analyzing the local 3D facial surface differences of the reconstructions and their corresponding ground truth. Identification success rate is obtained by comparing, based on correlation, Euclidean distance matrix (EDM) signatures of the reconstructed and the original 3D facial surfaces in the database. A subjective identification success rate is quantified based on face-pool tests. Finally a qualitative comparison is made between facial reconstructions of a real-case skull, based on two typical static face models and our statistical model, showing the shortcomings of current face models and the improved performance of the statistical model.

Sexual dimorphism in multiple aspects of 3D facial symmetry and asymmetry defined by spatially dense geometric morphometrics

Accurate measurement of facial sexual dimorphism is useful to understanding facial anatomy and specifically how faces influence, and have been influenced by, sexual selection. An important facial aspect is the display of bilateral symmetry, invoking the need to investigate aspects of symmetry and asymmetry separately when examining facial shape. Previous studies typically employed landmarks that provided only a sparse facial representation, where different landmark choices could lead to contrasting outcomes. Furthermore, sexual dimorphism is only tested as a difference of sample means, which is statistically the same as a difference in population location only. Within the framework of geometric morphometrics, we partition facial shape, represented in a spatially dense way, into patterns of symmetry and asymmetry, following a two-factor anova design. Subsequently, we investigate sexual dimorphism in symmetry and asymmetry patterns separately, and on multiple aspects, by examining (i) population location differences as well as differences in population variance-covariance; (ii) scale; and (iii) orientation. One important challenge in this approach is the proportionally high number of variables to observations necessitating the implementation of permutational and computationally feasible statistics. In a sample of gender-matched young adults (18-25 years) with self-reported European ancestry, we found greater variation in male faces than in women for all measurements. Statistically significant sexual dimorphism was found for the aspect of location in both symmetry and asymmetry (directional asymmetry), for the aspect of scale only in asymmetry (magnitude of fluctuating asymmetry) and, in contrast, for the aspect of orientation only in symmetry. Interesting interplays with hypotheses in evolutionary and developmental biology were observed, such as the selective nature of the force underpinning sexual dimorphism and the genetic independence of the structural patterns of fluctuating asymmetry. Additionally, insights into growth patterns of the soft tissue envelope of the face and underlying skull structure can also be obtained from the results.

Computerized craniofacial reconstruction using CT-derived implicit surface representations

In forensic craniofacial reconstruction, facial features of an unknown individual are estimated from an unidentified skull, based on a mixture of experimentally obtained guidelines on the relationship between soft tissues and the underlying skeleton. In this paper, we investigate the possibility of using full 3D cross-sectional CT images for establishing a reference database of densely sampled distances between the external surfaces of the skull and head for automated craniofacial reconstruction. For each CT image in the reference database, the hard tissue (skull) and soft tissue (head) volumes are automatically segmented and transformed into signed distance transform (sDT) images, representing for each voxel in this image the Euclidean distance to the closest point on the skull and head surface, respectively, distances being positive (negative) for voxels inside (outside) the skull/head. Multiple craniofacial reconstructions are obtained by first warping each reference skull sDT maps to the target skull sDT using a B-spline based free form deformation algorithm and subsequently applying these warps to the reference head sDT maps. A single reconstruction of the target head surface is defined as the zero level set of the arithmetic average of all warped reference head sDT maps, but other reconstructions are possible, biasing the result to subject specific attributes (age, BMI, gender). Both qualitative and quantitative tests (measuring the similarity between the 3D reconstructed and corresponding original head surface) on a small (N = 20) database are presented to proof the validity of the concept.

Evaluation of image features and search strategies for segmentation of bone structures in radiographs using Active Shape Models

In this paper, we evaluate various image features and different search strategies for fitting Active Shape Models (ASM) to bone object boundaries in digitized radiographs. The original ASM method iteratively refines the pose and shape parameters of the point distribution model driving the ASM by a least squares fit of the shape to update the target points at the estimated object boundary position, as determined by a suitable object boundary criterion. We propose an improved search procedure that is more robust against outlier configurations in the boundary target points by requiring subsequent shape changes to be smooth, which is imposed by a smoothness constraint on the displacement of neighbouring target points at each iteration and implemented by a minimal cost path approach. We compare the original ASM search method and our improved search algorithm with a third method that does not rely on iteratively refined target point positions, but instead optimizes a global Bayesian objective function derived from statistical a priori contour shape and image models. Extensive validation of these methods on a database containing more than 400 images of the femur, humerus and calcaneus using the manual expert segmentation as ground truth shows that our minimal cost path method is the most robust. We also evaluate various measures for capturing local image appearance around each boundary point and conclude that the Mahalanobis distance applied to normalized image intensity profiles extracted normal to the shape is the most suitable criterion among the tested ones for guiding the ASM optimization.

Effect of Lower Third Molar Segmentations on Automated Tooth Development Staging using a Convolutional Neural Network

Staging third molar development is commonly used for age estimation in subadults. Automated developmental stage allocation to the mandibular left third molar in panoramic radiographs has been examined in a pilot study. This method used an AlexNet Deep Convolutional Neural Network (CNN) approach to stage lower left third molars, which had been selected by manually drawn bounding boxes around them. This method (bounding box AlexNet = BA) still contained parts of surrounding structures which may have affected the automated stage allocation performance. We hypothesize that segmenting only the third molar could further improve the automated stage allocation performance. Therefore, the current study aimed to determine and validate the effect of lower third molar segmentations on automated tooth development staging. Retrospectively, 400 panoramic radiographs were collected, processed and segmented in three ways: bounding box (BB), rough (RS), and full (FS) tooth segmentation. A DenseNet201 CNN was used for automated stage allocation. Automated staging results were compared with reference stages - allocated by human observers - overall and per stage. FS rendered the best results with a stage allocation accuracy of 0.61, a mean absolute difference of 0.53 stages and a Cohen's linear κ of 0.84. Misallocated stages were mostly neighboring stages, and DenseNet201 rendered better results than AlexNet by increasing the percentage of correctly allocated stages by 3% (BA compared to BB). FS increased the percentage of correctly allocated stages by 7% compared to BB. In conclusion, full tooth segmentation and a DenseNet CNN optimize automated dental stage allocation for age estimation.