3D Segmentation Research Articles

Transformer based 3D tooth segmentation via point cloud region partition

Transformer based 3D tooth segmentation via point cloud region partition

Semi-supervised 3D retinal fluid segmentation via correlation mutual learning with global reasoning attention

Semi-supervised 3D retinal fluid segmentation via correlation mutual learning with global reasoning attention

Abstract 4141445: Long-Term Predictive Value of 4D Flow MRI in Bicuspid Aortic Valve Patients: A 10-Year Assessment for Aortic Surgery Risk

Introduction: Bicuspid aortic valve (BAV) is associated with progressive ascending aorta (AAo) dilation, often leading to aneurysms, dissections, and ruptures. Thus, current guidelines recommend preventive surgery for AAo dilation. Recent 4D flow MRI studies show that BAV morphology causes abnormal transvalvular flow patterns, increasing wall shear stress (WSS), a trigger of aortic growth. Further studies have delineated areas of abnormally high WSS by comparing to estimates of matched controls, and show promise in detecting risk for aortic growth. However, since the long-term prognostic significance of this marker is unclear, we aimed to quantify WSS in BAV patients to assess its value in predicting the need for aortic surgery up to 10 years post-4D flow MRI acquisition. Methods: BAV patients without prior surgical intervention scanned before April 1, 2014 were identified. Using medical records, patients were categorized as ‘operated’ if they underwent aortic surgery post-scan and ‘non-operated’ if they were surgery-free for at least 10 years post-scan. 4D flow MRIs were processed with an AI pipeline, including 3D segmentation of the aorta, followed by peak velocity (PV) and WSS quantification in the AAo (Fig. 1A-C). Patient-specific WSS heatmaps were computed relative to a map based on the WSS of 10 or more sex and age-matched controls. Relative areas of elevated WSS in the AAo were then calculated (Fig. 1D-F). Results: 115 patients were included, with 73 non-operated (age: 42.5±11.5y, 49M) and 42 operated patients (age: 53.5±12.1y, 34M). The mean baseline mid-AAo diameters for non-operated and operated patients were 3.8±0.6 cm and 4.1±0.5 cm, respectively. Among operated patients, the mean scan to surgery time was 5.7±3.3y. All three 4D flow metrics were significantly higher in operated compared to non-operated patients: PV: 2.6±0.6 vs. 1.7±0.4 m/s (p<0.01); WSS: 2.2±0.5 vs. 1.6±0.4 Pa (p<0.01); relative area of high WSS on heatmaps: 31±12 vs. 15±13% (p<0.01). Conclusions: 4D flow MRI parameters provide long-term predictive value in BAV patients, as elevated peak velocity, wall shear stress, and relative area of high wall shear stress on heatmaps are associated with subsequent aortic surgery within 10 years post-scan.

NIMG-05. ARTIFICIAL INTELLIGENCE-BASED RESPONSE ASSESSMENT FOR PATIENTS RECEIVING LASER INTERSTITIAL THERMAL THERAPY

Abstract BACKGROUND Laser interstitial thermal therapy (LITT) is a minimally invasive option for tissue diagnosis, cytoreduction, and rapid post-operative return to systemic therapies for patients with brain tumors or radiation necrosis. As ablated tissue remains in situ, post-LITT edema may be associated with apparent increased lesion size and transient clinical worsening, complicating assessment of progression. METHODS All patients receiving LITT at a single center for tumors or radiation necrosis from 2015 – 2023 with ≥ 9 months of MRI follow-up were retrospectively reviewed. A 3D U-net segmentation model implemented in nnU-Net was developed for automated segmentation of Contrast-enhancing Lesion Volume (CeLV) of LITT-treated lesions on T1-weighted post-contrast MR images. CeLVs were analyzed to establish volumetric post-LITT response assessments. RESULTS Across 384 unique MRI exams, 61LITT-treated lesions and 6 control cases of medically-managed radiation necrosis were analyzed. Automated segmentation was qualitatively accurate in 367/384 (95.6%) images. CeLV increased to a median of 68.3% (IQR 35.1 – 109.2%) from baseline at 1 – 3 months from LITT (P = 0.0012) and returned to baseline thereafter. Based on established criteria, volumetric disease progression was defined as lesion expansion 40% from volumetric nadir or baseline. Twenty-one of 56 (37.5%) patients experienced volumetric progression with a progression-free survival of 21.4 (6.0 – 93.4) months. Patients experiencing volumetric progression had lower mOS (17.3 vs 62.1 months, P = 0.0015). CONCLUSIONS We observed a stereotyped increase in CeLV at 1 – 3 months post-LITT, which resolved within 6 months of the procedure. Post-LITT disease response can be reliably and feasibly assessed with criteria that account for this transient CeLV expansion. Automated lesion segmentation may speed adoption of volumetric response assessment criteria to clinical practice.

Bessel beam optical coherence microscopy enables multiscale assessment of cerebrovascular network morphology and function

Understanding the morphology and function of large-scale cerebrovascular networks is crucial for studying brain health and disease. However, reconciling the demands for imaging on a broad scale with the precision of high-resolution volumetric microscopy has been a persistent challenge. In this study, we introduce Bessel beam optical coherence microscopy with an extended focus to capture the full cortical vascular hierarchy in mice over 1000 × 1000 × 360 μm3 field-of-view at capillary level resolution. The post-processing pipeline leverages a supervised deep learning approach for precise 3D segmentation of high-resolution angiograms, hence permitting reliable examination of microvascular structures at multiple spatial scales. Coupled with high-sensitivity Doppler optical coherence tomography, our method enables the computation of both axial and transverse blood velocity components as well as vessel-specific blood flow direction, facilitating a detailed assessment of morpho-functional characteristics across all vessel dimensions. Through graph-based analysis, we deliver insights into vascular connectivity, all the way from individual capillaries to broader network interactions, a task traditionally challenging for in vivo studies. The new imaging and analysis framework extends the frontiers of research into cerebrovascular function and neurovascular pathologies.

Estimation of outer-wall length in optimizing cochlear implantation in malformed inner ears

Estimation of cochlear length is gaining attention in the field of cochlear implants (CIs), mainly for selecting of CI electrode lengths. The currently available tools to estimate the cochlear duct length (CDL) are only valid for normal inner anatomy. However, inner ear malformation (IEM) types are associated with different degrees of cystic apices, limiting the application of CDL equations of normal anatomy inner ear. Therefore, this study aimed to understand the degree to which the outer wall (OW) is observed in different malformation types and to formulate mathematical equations to estimate the OW length (OWL) from cochlear parameters, namely the basal turn diameter (A-value) and width (B-value). Three-dimensional (3D) segmentation of promontory and fluid parts of the inner ear was performed to understand the extent to which the OW is visible to measure the OWL manually. Enlarged vestibular aqueduct syndrome (EVAS) was diagnosed in 37 ears, which consistently showed the extent of the OW to an angular depth of 540°, beyond which the cystic apex starts. Incomplete partition (IP) type I was observed in 30 ears, with the OW extending to only 360° of angular depth. IP type II was observed in 35 ears, with the OW extending to 450° of angular depth. IP type III was identified in 24 ears, with the OW observed for 540° of angular depth. Cavity-type malformations were observed in 36 ears, and circumference was measured in the axial view. A strong positive linear correlation was observed between the manually measured OWL and cochlear parameters for all malformation types analyzed. A multiple linear regression model was applied to formulate mathematical equations, which was further used to create a software application for estimating OWLs in IEM types, using cochlear parameters as inputs.

Leaf Identification in High-Density LiDAR-RGB Data

Abstract. Leaf detection through automated segmentation of 3D data is becoming a crucial technique in many applications of digital agriculture. Some 3D segmentation techniques that can be mentioned are based on normal differences and median normalised vector growth. However, applying these approaches to high canopy density data remains challenging. In this study, we propose a processing flow for leaf detection in high canopy density LiDAR-RGB point clouds. First, a noise removal technique inspired by Moving Least Squares (MLS) was applied to the LiDAR point cloud, and a RGB colour was assigned to each point by combining computer vision and photogrammetric methods. Moreover, once the data were suitable for leaf detection, the branches were filtered using the Statistical Outlier Removal (SOR) filter based on an analysis of the statistical behaviour of the neighbourhood. Afterwards, an unsupervised DBSCAN (Density-Based Clustering Non-Parametric Algorithm) method was used to segment similar points. Finally, the points within each cluster were identified as leaf or non-leaf using the RGB values implemented by our method; ground points were filtered out using a maximum height threshold. As a result, the leaf, non-leaf, and ground point identifiers were correct in 98.9% of cases, with the branch filtering technique SOR proving effectiveness in removing branches with low information loss and without additional complex point densification steps in reconstruction. This SOR-based solution overcomes major challenges in semantic segmentation (leaves and branches) in high-density data and potentially contributes to precision agriculture.

A flexible 2.5D medical image segmentation approach with in-slice and cross-slice attention

Deep learning has become the de facto method for medical image segmentation, with 3D segmentation models excelling in capturing complex 3D structures and 2D models offering high computational efficiency. However, segmenting 2.5D images, characterized by high in-plane resolution but lower through-plane resolution, presents significant challenges. While applying 2D models to individual slices of a 2.5D image is feasible, it fails to capture the spatial relationships between slices. On the other hand, 3D models face challenges such as resolution inconsistencies in 2.5D images, along with computational complexity and susceptibility to overfitting when trained with limited data. In this context, 2.5D models, which capture inter-slice correlations using only 2D neural networks, emerge as a promising solution due to their reduced computational demand and simplicity in implementation. In this paper, we introduce CSA-Net, a flexible 2.5D segmentation model capable of processing 2.5D images with an arbitrary number of slices. CSA-Net features an innovative Cross-Slice Attention (CSA) module that effectively captures 3D spatial information by learning long-range dependencies between the center slice (for segmentation) and its neighboring slices. Moreover, CSA-Net utilizes the self-attention mechanism to learn correlations among pixels within the center slice. We evaluated CSA-Net on three 2.5D segmentation tasks: (1) multi-class brain MR image segmentation, (2) binary prostate MR image segmentation, and (3) multi-class prostate MR image segmentation. CSA-Net outperformed leading 2D, 2.5D, and 3D segmentation methods across all three tasks, achieving average Dice coefficients and HD95 values of 0.897 and 1.40 mm for the brain dataset, 0.921 and 1.06 mm for the prostate dataset, and 0.659 and 2.70 mm for the ProstateX dataset, demonstrating its efficacy and superiority. Our code is publicly available at: https://github.com/mirthAI/CSA-Net.

Detection of clavicle fracture after martial arts training based on 3D segmentation and multi-perspective fusion

To improve the diagnosis efficiency of clavicle fractures, a Diagnosis Model of Clavicle Fracture Based on 3D Segmentation and Multi Perspective Fusion (3Ds MPF) is proposed in the experiment. By performing 3D segmentation on clavicle images, key layer images are extracted, and multi-perspective image data are fused and output to achieve accurate classification and diagnosis results. The results show that on the relevant dataset, the loss function value of the constructed model is less than 0.01, and the calculation accuracy is as high as 0.999. The average classification accuracy in actual use exceeds 90%, and it can capture the majority of fracture features in clavicle computed tomography images. The above results indicate that the model constructed by the research institute can significantly improve the judgment efficiency of computed tomography images, which helps to enhance the detection and diagnosis of clinical fracture conditions. The experimental results have helped improve the clinical diagnostic efficiency of fracture detection work.

Temporomandibular joint CBCT image segmentation via multi-view ensemble learning network.

Accurate segmentation of the temporomandibular joint (TMJ) from cone beam CT (CBCT) images holds significant clinical value for diagnosing temporomandibular joint osteoarthrosis (TMJOA) and related conditions. Convolutional neural network-based medical image segmentation methods have achieved state-of-the-art performance in various segmentation tasks. However, 3D medical images segmentation requires substantial global context and rich spatial semantic information, demanding much more GPU memory and computational resources. To address these challenges in 3D medical image segmentation, we propose a novel network- the MVEL-Net (Multi-view Ensemble Learning Network) for TMJ CBCT image segmentation. By resampling images along three dimensions, we generate multiple weak learners with different spatial semantic information. A subsequent strong learning network effectively integrates the outputs from these weak learners to achieve more accurate segmentation results. We evaluated our network model using a clinical dataset comprising 88 subjects with TMJ CBCT images. The average Dice similarity coefficient (DSC) was 0.9817 ± 0.0049, the average surface distance was 0.0540 ± 0.0179mm, and the 95% Hausdorff distance was 0.1743 ± 0.0550mm. Our proposed MVEL-Net demonstrates excellent segmentation performance on TMJ from CBCT images, while using fewer GPU memory resources compared to other 3D networks. The effectiveness of this method in capturing spatial context could be leveraged for tasks like organ segmentation from volumetric scans. This may facilitate wider adoption of AI-based solutions for automated analysis of 3D medical images.

Deep learning enables the fully automated quantification of high-resolution cardiac CT images: human in loop for model development

Abstract Introduction Cardiac computed tomography (CCT) images provide vast spatial and temporal information. Segmentation of the structure of interest is necessary to provide quantitative information. However, manual segmentation of cardiac substructures is highly time-consuming. Objective This study aims to develop a robust deep-learning segmentation model for cardiac substructure segmentation in 3D CCT images. Method We initially developed a 3D U-Net architecture, incorporating extensive augmentation, for the segmentation of the Left Ventricle Myocardium (LVM), the blood cavities of the Left and Right Ventricles (LV, RV), and the Left and Right Atria (LA, RA). This model utilized CT images from 20 patients who had available manual segmentation. Subsequently, we applied this initial model to datasets from two centers, including different generations of scanners. We then visually inspected the segmentation, selected the highest-quality segmentations, and modify it if necessary, and excluded the low-quality data. We continued fine-tuning the model with the new dataset until we reached a count of 456 patients. As there were multiple images from one patient during this model's development, we generated datasets to develop a new model using the entire new dataset to avoid any data leakage in the reported metrics. The model was developed using data from Center 1 (5-fold cross-validation) and then evaluated on an external test set from Center 2 (126 patients) and also 4D images from Center 1. Various metrics were calculated to assess the model's performance. Results In the internal test set, utilizing 5-fold CV, the model demonstrated high segmentation accuracy, achieving an overall Dice coefficient of 0.98, with individual scores of 0.96 for LVM, 0.98 for LV, 0.98 for RV, 0.99 for RA, and 0.98 for LA. We assessed the automatic quantification capability of the segmentation model from 256 patients, resulting in an R2 value greater than 0.92 for both LV end-diastolic and systolic volumes. In contrast, the model maintained robust performance on the external dataset, achieving an overall Dice coefficient of 0.96, with individual scores of 0.95 for LVM, 0.98 for LV, 0.97 for RV, 0.98 for RA, and 0.94 for LA. Conclusion We provided a large cohort of ground truth segmentations for different cardiac CT substructures using a human-in-the-loop strategy. Subsequently, we developed a deep learning model for highly accurate automated segmentation of these substructures. This model allows for fast and fully automatic segmentation and quantification of 3D and 4D cardiac CT images, from which various quantitative metrics can be extracted from different cardiac substructures.

Fast‐MedNeXt: Accelerating the MedNeXt Architecture to Improve Brain Tumour Segmentation Efficiency

ABSTRACTWith the rapid development of medical imaging technology, 3D image segmentation technology has gradually become a mainstream method, especially in brain tumour detection and diagnosis showing its unique advantages. The technique makes full use of 3D spatial information to locate and analyze tumours more accurately, thus playing an important role in improving diagnostic accuracy, optimising treatment planning and promoting research. However, it also suffers from significant computational expenditure and delayed processing pace. In this paper, we propose an innovative optimisation scheme to address this problem. We thoroughly investigate the MedNeXt network and propose Fast‐MedNeXt, which aims to increase the processing speed while maintaining accuracy. First, we introduce the partial convolution (PConv) technique, which replaces the deep convolutional layers in the network. This improvement effectively reduces computation and memory requirements while maintaining efficient feature extraction. Second, based on PConv, we propose PConv‐Down and PConv‐Up modules, which are applied to the up‐sampling and down‐sampling modules to further optimise the network structure and improve efficiency. To confirm the efficacy of the approach, we carried out a sequence of tests in the multimodal brain tumour segmentation challenge 2021 (BraTS2021). By comparing with the MedNeXt series network, the Fast‐MedNeXt reduced the latency by 22.1%, 20.5%, 15.8%, and 11.4% respectively, while the average accuracy also increased by 0.475% and 0.2% respectively. These significant performance improvements demonstrate the effectiveness of Fast‐MedNeXt in 3D medical image segmentation tasks and provide a new and more efficient solution for the field.

RMCNet: A Liver Cancer Segmentation Network Based on 3D Multi-Scale Convolution, Attention, and Residual Path

Abdominal CT images are important clues for diagnosing liver cancer lesions. However, liver cancer presents challenges such as significant differences in tumor size, shape, and location, which can affect segmentation accuracy. To address these challenges, we propose an end-to-end 3D segmentation algorithm, RMCNet. In the shallow encoding part of RMCNet, we incorporated a 3D multiscale convolution (3D-Multiscale Convolution) module to more effectively extract tumors of varying sizes. Moreover, the convolutional block attention module (CBAM) is used in the encoding part to help the model focus on both the shape and location of tumors. Additionally, a residual path is introduced in each encoding layer to further enrich the extracted feature maps. Our method achieved DSC scores of 76.56% and 72.96%, JCC scores of 75.82% and 71.25%, HD values of 11.07 mm and 17.06 mm, and ASD values of 2.54 mm and 10.51 mm on the MICCAI 2017 Liver Tumor Segmentation public dataset and the 3Dircadb-01 public dataset, respectively. Compared to other methods, RMCNet demonstrates superior segmentation performance and stronger generalization capability.

Sparse Annotation is Sufficient for Bootstrapping Dense Segmentation.

Producing dense 3D reconstructions from biological imaging data is a challenging instance segmentation task that requires significant ground-truth training data for effective and accurate deep learning-based models. Generating training data requires intense human effort to annotate each instance of an object across serial section images. Our focus is on the especially complicated brain neuropil, comprising an extensive interdigitation of dendritic, axonal, and glial processes visualized through serial section electron microscopy. We developed a novel deep learning-based method to generate dense 3D segmentations rapidly from sparse 2D annotations of a few objects on single sections. Models trained on the rapidly generated segmentations achieved similar accuracy as those trained on expert dense ground-truth annotations. Human time to generate annotations was reduced by three orders of magnitude and could be produced by non-expert annotators. This capability will democratize generation of training data for large image volumes needed to achieve brain circuits and measures of circuit strengths.

Adversarial Unsupervised Domain Adaptation for 3D Semantic Segmentation with 2D Image Fusion of Dense Depth

AbstractUnsupervised domain adaptation (UDA) is increasingly used for 3D point cloud semantic segmentation tasks due to its ability to address the issue of missing labels for new domains. However, most existing unsupervised domain adaptation methods focus only on uni‐modal data and are rarely applied to multi‐modal data. Therefore, we propose a cross‐modal UDA on multi‐modal datasets that contain 3D point clouds and 2D images for 3D Semantic Segmentation. Specifically, we first propose a Dual discriminator‐based Domain Adaptation (Dd‐bDA) module to enhance the adaptability of different domains. Second, given that the robustness of depth information to domain shifts can provide more details for semantic segmentation, we further employ a Dense depth Feature Fusion (DdFF) module to extract image features with rich depth cues. We evaluate our model in four unsupervised domain adaptation scenarios, i.e., dataset‐to‐dataset (A2D2 → SemanticKITTI), Day‐to‐Night, country‐to‐country (USA → Singapore), and synthetic‐to‐real (VirtualKITTI → SemanticKITTI). In all settings, the experimental results achieve significant improvements and surpass state‐of‐the‐art models.

Point cloud segmentation method based on an image mask and its application verification

Abstract Accurately perceiving three-dimensional (3D) environments or objects is crucial for the advancement of artificial intelligence (AI) interaction technologies. Currently, various types of sensors are employed to obtain point cloud data for 3D object detection or segmentation tasks. While this multi-sensor approach provides more precise 3D data than monocular or stereo cameras, it is also more expensive. The advent of RGB-D cameras, which provide both RGB images and depth information, addresses this issue.
In this study, we propose a point cloud segmentation method based on image masks. By using an RGB-D camera to capture color and depth images, we generate image masks through object recognition and segmentation. Given the mapping relationship between RGB image pixels and point clouds, these image masks can be further used to extract the point cloud data of the target objects. The experimental results revealed that the average accuracy of target segmentation was 84.78%, which was close to that of PointNet++. Compared with three traditional segmentation algorithms, the accuracy was improved by nearly 23.97%. The running time of our algorithm is reduced by 95.76% compared to the PointNet++ algorithm, which has the longest running time; and by 15.65% compared to the LCCP algorithm, which has the shortest running time among traditional methods. Compared with PointNet++, the segmentation accuracy was improved. This method addressed the issues of low robustness and excessive reliance on manual feature extraction in traditional point cloud segmentation methods, providing valuable support and reference for the accurate segmentation of 3D point clouds.

Automated volumetric analysis of the inner ear fluid space from hydrops magnetic resonance imaging using 3D neural networks

Due to the development of magnetic resonance (MR) imaging processing technology, image-based identification of endolymphatic hydrops (EH) has played an important role in understanding inner ear illnesses, such as Meniere’s disease or fluctuating sensorineural hearing loss. We segmented the inner ear, consisting of the cochlea, vestibule, and semicircular canals, using a 3D-based deep neural network model for accurate and automated EH volume ratio calculations. We built a dataset of MR cisternography (MRC) and HYDROPS-Mi2 stacks labeled with the segmentation of the perilymph fluid space and endolymph fluid space of the inner ear to devise a 3D segmentation deep neural network model. End-to-end learning was used to segment the perilymph fluid and the endolymph fluid spaces simultaneously using aligned pair data of the MRC and HYDROPS-Mi2 stacks. Consequently, the segmentation performance of the total fluid space and endolymph fluid space had Dice similarity coefficients of 0.9574 and 0.9186, respectively. In addition, the EH volume ratio calculated by experienced otologists and the EH volume ratio value predicted by the proposed deep learning model showed high agreement according to the interclass correlation coefficient (ICC) and Bland–Altman plot analysis.

AI-based 3D analysis of retinal vasculature associated with retinal diseases using OCT angiography

Retinal vasculature is the only vascular system in the human body that can be observed in a non-invasive manner, with a phenotype associated with a wide range of ocular, cerebral, and cardiovascular diseases. OCT and OCT angiography (OCTA) provide powerful imaging methods to visualize the three-dimensional morphological and functional information of the retina. In this study, based on OCT and OCTA multimodal inputs, a multitask convolutional neural network model was built to realize 3D segmentation of retinal blood vessels and disease classification for different retinal diseases, overcoming the limitations of existing methods that can only perform 2D analysis of OCTA. Two hundred thirty sets of OCT and OCTA data from 109 patients, including 138,000 cross-sectional images in normal and diseased eyes (age-related macular degeneration, retinal vein occlusion, and central serous chorioretinopathy), were collected from four commercial OCT systems for model training, validation, and testing. Experimental results verified that the proposed method was able to achieve a DICE coefficient of 0.956 for 3D segmentation of blood vessels and an accuracy of 91.49% for disease classification, and further enabled us to evaluate the 3D reconstruction of retinal vessels, explore the interlayer connections of superficial and deep vasculatures, and reveal the 3D quantitative vessel characteristics in different retinal diseases.

Weakly supervised 3D point cloud semantic segmentation for architectural heritage using teacher-guided consistency and contrast learning

Point cloud semantic segmentation is significant for managing and protecting architectural heritage. Currently, fully supervised methods require a large amount of annotated data, while weakly supervised methods are difficult to transfer directly to architectural heritage. This paper proposes an end-to-end teacher-guided consistency and contrastive learning weakly supervised (TCCWS) framework for architectural heritage point cloud semantic segmentation, which can fully utilize limited labeled points to train network. Specifically, a teacher-student framework is designed to generate pseudo labels and a pseudo label dividing module is proposed to distinguish reliable and ambiguous point sets. Based on it, a consistency and contrastive learning strategy is designed to fully utilize supervision signals to learn the features of point clouds. The framework is tested on the ArCH dataset and self-collected point cloud, which demonstrates that the proposed method can achieve effective semantic segmentation of architectural heritage using only 0.1 % of annotated points.

Stone instance segmentation of rubble masonry based on laser scanning point clouds

Laser scanning allows for objective area-based analysis of water dam structures to enable targeted interventions in case of displacements, requiring comparison of the same areas over different epochs. This comparison improves if identical areas, e.g. stones, from multiple epochs can be automatically derived in advance. We present an instance segmentation algorithm based on a 3D point cloud with reflected intensity information to identify individual stones in rubble masonry dams that can be used within deformation monitoring. The algorithm uses initial k-means classification followed by image-based processing steps, resulting in a segmented 3D point cloud with individual stones. It is evaluated against a manually labeled reference and analyzed on terrestrial laser scanning (TLS) and UAV-based data sets. Correctly identified stones are 89.92% for TLS and 90.82% for UAV data. Additionally, stone centroids and shapes were evaluated without significant deviation. A first outlook on deformation analysis was provided using ICP matching of stones from two simulated epochs.