Segmentation Errors Research Articles

Interobserver regional variability of left ventricular endocardium and epicardium delineation in echocardiography imaging

Abstract Introduction Accurate delineation of the left ventricle (LV) endocardium and epicardium remains a significant challenge in echocardiography due to low image quality and tissue contrast, leading to a compromised reproducibility of LV border tracing [1]. Whereas the global variability in echocardiography segmentation is extensively discussed in the literature [2], there is still a substantial gap in understanding the distribution of border delineation error across the different regions of LV. Purpose This study aims to extensively analyse discrepancies in the echocardiography delineation of LV endocardial and epicardial borders at global and regional levels by considering a large cohort of annotation experts. Methods LV endocardial and epicardial borders were independently delineated by 35 sonographers from multiple institutions in apical 4-chamber (A4C) and 2-chamber (A2C) echocardiography sequences at end-diastole (ED) and end-systole (ES) in a cohort of 10 patients, totalling 80 annotations per observer. The ES and ED frames were preselected, and their ground truth was determined by a clinical committee. LV sections were obtained via standard Simpson’s rule [3], yielding 82 reference points along the epicardium and endocardium, grouped into 4 regions (septal, lateral, anterior, inferior) across 3 levels (basal, mid, apical) plus the apex, resulting in 13 segments. Global segmentation error was quantified using the Dice Similarity Coefficient (DSC) and Hausdorff 95% distance (HD95). Regional error was estimated as the minimum Euclidean distance from each of the 82 ground truth reference points to the segmentation border, averaged across LV segments and patients. Statistical significance was assessed via paired t-test. Results The average DSC was 0.87±0.03 and 0.92±0.02, and the average HD95 was 6.0±1.4 and 5.8±1.1 mm for endocardium and epicardium, respectively (Figure 1). Segmentation error was significantly lower at ED compared to ES (p=0.003) and in A4C compared to A2C (p=0.001). In line with the latter, the error in the septal region was significantly the lowest (Figure 2). A positive error trend base-to-apex is observed, with the error significantly the smallest at the basal level. Errors were more prominent in the apical lateral endocardial segment in A4C and in the apical anterior segment in A2C. These errors can be explained by the higher LV curvature in these segments [4]. Epicardium apex error could have been slightly underestimated due to some of the contours falling outside of the field of view in this region. Conclusion This work introduces a reproducible method to analyse global and regional variability in LV border tracing in echocardiography. The study identifies compromised segments that can be targeted via enhanced training and standardised protocols to effectively reduce interobserver variability, leading to consistent and accurate interpretations, ultimately benefiting patient care.

A Named Entity Recognition Model for Chinese Electricity Violation Descriptions Based on Word-Character Fusion and Multi-Head Attention Mechanisms

Due to the complexity and technicality of named entity recognition (NER) in the power grid field, existing methods are ineffective at identifying specialized terms in power grid operation record texts. Therefore, this paper proposes a Chinese power violation description entity recognition model based on word-character fusion and multi-head attention mechanisms. The model first utilizes a collected power grid domain corpus to train a Word2Vec model, which produces static word vector representations. These static word vectors are then integrated with the dynamic character vector features of the input text generated by the BERT model, thereby mitigating the impact of segmentation errors on the NER model and enhancing the model’s ability to identify entity boundaries. The combined vectors are subsequently input into a BiGRU model for learning contextual features. The output from the BiGRU layer is then passed to an attention mechanism layer to obtain enhanced semantic features, which highlight key semantics and improve the model’s contextual understanding ability. Finally, the CRF layer decodes the output to generate the globally optimal label sequence with the highest probability. Experimental results on the constructed power grid field operation violation description dataset demonstrate that the proposed NER model outperforms the traditional BERT-BiLSTM-CRF model, with an average improvement of 1.58% in precision, recall, and F1-score. This demonstrates the effectiveness of the model design and further enhances the accuracy of entity recognition in the power grid domain.

An Improved Tree Crown Delineation Method Based on a Gradient Feature-Driven Expansion Process Using Airborne LiDAR Data

Accurate individual tree crown delineation (ITCD), which can be used to estimate various forest parameters such as biomass, stem density, and carbon storage, stands as an essential component of precision forestry. Currently, raster data such as the canopy height model derived from airborne light detection and ranging (LiDAR) data have been widely used in large-scale ITCD. However, the accuracy of current existing algorithms is limited due to the influence of understory vegetation and variations in tree crown geometry (e.g., the delineated crown boundaries consistently extend beyond their actual boundaries). In this study, we achieved more accurate crown delineation results based on an expansion process. First, the initial crown boundaries were extracted through watershed segmentation. Then, a “from the inside out” expansion process was guided by a novel gradient feature to obtain accurate crown delineation results across different forest conditions. Results show that our method produced much better performance (~75% matched on average) than other commonly used methods across all test forest plots. The erroneous situation of “match but over-grow” is significantly reduced, regardless of forest conditions. Compared to other methods, our method demonstrates a notable increase in the precisely matched rate across different plot types, with an average increase of 25% in broadleaf plots, 18% in coniferous plots, 23% in mixed plots, 15% in high-density plots, and 32% in medium-density plots, without increasing over- and under- segmentation errors. Our method demonstrates potential applicability across various forest conditions, facilitating future large-scale ITCD tasks and precision forestry applications.

Semi-Automatic Refinement of Myocardial Segmentations for Better LVNC Detection.

Background: Accurate segmentation of the left ventricular myocardium in cardiac MRI is essential for developing reliable deep learning models to diagnose left ventricular non-compaction cardiomyopathy (LVNC). This work focuses on improving the segmentation database used to train these models, enhancing the quality of myocardial segmentation for more precise model training. Methods: We present a semi-automatic framework that refines segmentations through three fundamental approaches: (1) combining neural network outputs with expert-driven corrections, (2) implementing a blob-selection method to correct segmentation errors and neural network hallucinations, and (3) employing a cross-validation process using the baseline U-Net model. Results: Applied to datasets from three hospitals, these methods demonstrate improved segmentation accuracy, with the blob-selection technique boosting the Dice coefficient for the Trabecular Zone by up to 0.06 in certain populations. Conclusions: Our approach enhances the dataset's quality, providing a more robust foundation for future LVNC diagnostic models.

A prospectively deployed deep learning-enabled automated quality assurance tool for oncological palliative spine radiation therapy.

Palliative spine radiation therapy is prone to treatment at the wrong anatomic level. We developed a fully automated deep learning-based spine-targeting quality assurance system (DL-SpiQA) for detecting treatment at the wrong anatomic level. DL-SpiQA was evaluated based on retrospective testing of spine radiation therapy treatments and prospective clinical deployment. The DL-SpiQA workflow involves auto-segmentation and labelling of all vertebral volumes on CT imaging using TotalSegmentator, an open-source deep learning algorithm based on nnU-Net, calculation of the radiation dose to each vertebra, and flagging and categorisation of potential treatments at the wrong anatomic level with automated email reports sent to involved radiation therapy personnel. We developed the DL-SpiQA tool based on retrospective clinical data from patients treated with palliative spine radiation therapy from sites included in the multicentre hospital network between Feb 12, 2014, and Nov 15, 2022. We used historic cases of patients who had a near-miss (ie, wrong-anatomic-level errors caught before the patient was treated) or had received wrong-anatomic-level treatment to test whether the tool could identify known errors successfully. We then used the tool prospectively over 15 months (April 24, 2023, to July 22, 2024) to evaluate any new spine radiation therapy treatment plan created for a patient, looking for any targeting errors, and dose and volume discrepancies. An email report was circulated with all the radiation therapy personnel; if any errors were found, these were highlighted and each error was defined. The tool was internally validated. All cases flagged by DL-SpiQA for both the retrospective and prospective studies were manually reviewed for dosimetric targeting, variant spine anatomy or spinal anomalies, and artificial intelligence (AI) segmentation errors. DL-SpiQA was further validated based on false positive and negative rates estimated from the retrospective results. DL-SpiQA was first tested retrospectively on 513 patients with segmentation of 10 106 vertebrae. The system raised flags for ten dose discrepancies, 49 normal anatomic variants, 49 cases with implants or other anomalies, and 20 segmentation errors (4% false positive rate). DL-SpiQA caught one historic treatment at the wrong anatomic level and three near-misses. DL-SpiQA was then prospectively deployed, reviewing 520 cases and identifying six documentation errors, which triggered detailed review by clinicians, and 43 additional cases, which confirmed clinical knowledge of variant anatomy. In all detected cases (ie, 49 of 520 cases in total), the appropriate personnel were alerted. A false negative rate of 0·03% is estimated based on the 4% AI segmentation error rate and the frequency of reported spine radiation therapy errors. The low false positive rate, the low false negative rate, and the high accuracy in flagging errors show that DL-SpiQA is an effective, AI-driven, automated quality assurance tool that could be used to identify anatomic spine variants and errors in targeting at the anatomic level. The tool could therefore help improve the safety of spine radiotherapy. Further external validation and tailoring is needed. None.

ImmuNet: a segmentation-free machine learning pipeline for immune landscape phenotyping in tumors by multiplex imaging.

Tissue specimens taken from primary tumors or metastases contain important information for diagnosis and treatment of cancer patients. Multiplex imaging allows in situ visualization of heterogeneous cell populations, such as immune cells, in tissue samples. Most image processing pipelines first segment cell boundaries and then measure marker expression to assign cell phenotypes. In dense tissue environments, this segmentation-first approach can be inaccurate due to segmentation errors or overlapping cells. Here, we introduce the machine-learning pipeline "ImmuNet", which identifies positions and phenotypes of cells without segmenting them. ImmuNet is easy to train: human annotators only need to click on an immune cell and score its expression of each marker-drawing a full cell outline is not required. We trained and evaluated ImmuNet on multiplex images from human tonsil, lung cancer, prostate cancer, melanoma, and bladder cancer tissue samples and found it to consistently achieve error rates below 5%-10% across tissue types, cell types, and tissue densities, outperforming a segmentation-based baseline method. Furthermore, we externally validate ImmuNet results by comparing them to flow cytometric cell count measurements from the same tissue. In summary, ImmuNet is an effective, simpler alternative to segmentation-based approaches when only cell positions and phenotypes, but not their shapes, are required for downstream analyses. Thus, ImmuNet helps researchers to analyze cell positions in multiplex tissue images more easily and accurately.

Frequency of segmental sound errors among intermediate and advanced Korean learners from Myanmar

Frequency of segmental sound errors among intermediate and advanced Korean learners from Myanmar

Tone-vowel interaction and co-articulation in Cantonese speakers with apraxia of speech and co-existing aphasia: a preliminary study

ABSTRACT Background Speakers with apraxia of speech (AOS) usually produce segmental and prosodic errors that influence their speech intelligibility. Literature on AOS in tonal language speakers is sparse compared to that in non-tonal language. Also, the existing research often made static, isolated analyses, leaving the production and co-articulation between segments and suprasegmental entities in tonal languages under-investigated. Aims This preliminary study aims to fulfill the aforementioned research gaps by investigating vowel-tone interaction and tonal and vocalic co-articulation in Cantonese post-stroke speakers with AOS. Methods Five Cantonese adults with AOS post-stroke, five adults without AOS post-stroke, and five healthy controls performed the Tone Sequencing Task (TST), a task adapted from oral diadochokinetic tasks that required five rapid repetitions of 3-syllable items formed by three different Cantonese vowels and three different Cantonese tones. The quality of vowels was indexed by midpoint formant values and euclidean distances between the vowels. Within-speaker variation was assessed by coefficients of variance. Co-articulation was indexed by onset and offset formant or f0 values. The effects of the participant groups, the positions of tone-syllable in the TST stimuli, and the tones carried by vowels/carrying vowels were evaluated with linear mixed effect models. Results Cantonese-speaking adults with AOS had difficulty in producing distinctive vowels and tones. They also showed large within-speaker variation in vowel production but reduced tone contrast, especially at the final positions. The disrupted anticipatory co-articulation between vowels as well as between tones further suggested that the speakers with AOS could not sequence segments and suprasegments simultaneously.

Combined Layer-by-Layer Segmentation Level Set Model Applied to Dental CBCT Image Segmentation

Cone beam computed tomography (CBCT) image technology is widely used in the oral healthcare service industry to Detect the alignment, shape, position, and orientation of the patient's teeth, and tooth segmentation based on the level set model is an important step in the reconstruction and visualization of the tooth's three-dimensional structure. In response to the error accumulation problem that occurs in the conventional layer-by-layer segmentation level set model, this paper proposes a combined layer-by-layer segmentation level set model, which reduces the number of segmentation error transfers, and at the same time reduces the amount of segmentation error accumulation. Two consecutive sets of level set function iterations are designed to lead the curve from the inside of the tooth to the tooth edge and perform accurate tooth contour segmentation. The first set moves the curve from the inside of the tooth to the vicinity of the tooth contour to avoid over-segmentation, and the second set moves the curve from the vicinity of the tooth contour to the edge of the tooth contour for further optimization. Ten CBCT images were randomly selected to test this model, and the experiments showed that the combined layer-by-layer segmentation level-set model was superior to the conventional layer-by-layer segmentation level-set model, with the VOE value reduced by 12.44%, and the results of tooth segmentation were more accurate and better displayed; the time spent was shorter, but not significant; and good segmentation accuracy could be achieved for various classes of teeth.

Multi-scale adaptive YOLO for instance segmentation of grape pedicels

To address the presence of misdetections, missed detections, and segmentation errors that easily occur in existing lightweight models when segmenting grape pedicels in vineyards with complex backgrounds, variable lighting conditions, and varying pedicel scales, this paper proposes a lightweight grape pedicel instance segmentation model based on an improved YOLOv8. The model is named Multi-Scale Adaptive YOLO (MSA-YOLO). Firstly, this paper proposes a Multi-scale Attention Mixing Head (MAMH), which allows for a wider network input to increase the scale range. It establishes high-level semantic associations between different scales and achieves output network compression through the Multi-scale Interactive-attention Module (MIM) and Cross-channel Mixing Module (CMM). Second, the paper presents the Multi-channel Enhanced Neck (MEN), which includes two main improvements: (1) adding a feature channel specifically for small targets in the Neck for a larger scale feature channel range; and (2) the Balanced Down-sampling Module (BDM) to retain more fine-grained features during the fusion of lower-level feature channels with higher-level feature channels. Finally, the paper presents the Star Fusion Module (SFM), which enhances the model’s ability to capture multi-scale features by integrating different subspace features in the backbone and increasing the diversity of features. The proposed MSA-YOLO model, with a significantly reduced GFLOPs of just 9.7 and a parameter count of only 2.3M, demonstrates exceptional performance on our self-constructed grape dataset. The model achieves AP@50 values of 99.3% for object detection and 99.5% for instance segmentation tasks. Compared to existing lightweight instance segmentation models, MSA-YOLO achieves better performance in segmenting grape pedicels under complex scenes.

RSANet: Relative-sequence quality assessment network for gait recognition in the wild

Gait recognition in the wild has received increasing attention since the gait pattern is hard to disguise and can be captured in a long distance. However, due to occlusions and segmentation errors, low-quality silhouettes are common and inevitable. To mitigate this low-quality problem, some prior arts propose absolute-single quality assessment models. Although these methods obtain a good performance, they only focus on the silhouette quality of a single frame, lacking consideration of the variation state of the entire sequence. In this paper, we propose a Relative-Sequence Quality Assessment Network, named RSANet. It uses the Average Feature Similarity Module (AFSM) to evaluate silhouette quality by calculating the similarity between one silhouette and all other silhouettes in the same silhouette sequence. The silhouette quality is based on the sequence, reflecting a relative quality. Furthermore, RSANet uses Multi-Temporal-Receptive-Field Residual Blocks (MTB) to extend temporal receptive fields without parameter increases. It achieves a Rank-1 accuracy of 75.2% on Gait3D, 81.8% on GREW, and 77.6% on BUAA-Duke-Gait datasets respectively. The code is available at https://github.com/PGZ-Sleepy/RSANet.

AUTOMATIC ASSESSMENT OF OCT SKIN THICKNESS IN MICE BASED ON TRANSFER LEARNING AND ATTENTION MECHANISMS

Optical coherence tomography (OCT) is characterized by high resolution and noninvasiveness; thus, it has been widely used to analyze skin tissues in recent years. Previous studies have evaluated skin OCT images using traditional algorithms with low accuracy for complex tissue structure images. Although a few studies have used deep learning methods to assess tissue structure in OCT images, they lack quantitative assessment of deeper skin tissue thickness and are limited to the epidermal layer. Thus, in the present study, we proposed an automated segmentation and quantitative evaluation method. The skin OCT images were first pre-processed, and the attention mechanism was added to U-Net based on transfer learning to segment the images and quantify the thickness of mouse skin tissue structure. The results showed that U-Net combined with the coordinate attention (CA) mechanism had better segmentation performance with 93.94% mean intersection of union (MIoU) value and 96.99% Dice similarity coefficient; the segmentation errors were 0.6 μm, 2.2 μm, 3.8 μm, and 6.0 μm for the epidermis layer, subcutaneous fat layer, muscle fascia layer, and the overall skin tissue structure of mice, respectively. The overall skin tissue thickness of the four mice were 235 ± 20 μm, 264 ± 42 μm, 275 ± 40 μm, and 774 ± 91 μm, respectively. The present study provides a rapid and accurate method for the automated measurement of skin tissue thickness.

Global Neuron Shape Reasoning with Point Affinity Transformers.

Connectomics is a subfield of neuroscience that aims to map the brain's intricate wiring diagram. Accurate neuron segmentation from microscopy volumes is essential for automating connectome reconstruction. However, current state-of-the-art algorithms use image-based convolutional neural networks that are limited to local neuron shape context. Thus, we introduce a new framework that reasons over global neuron shape with a novel point affinity transformer. Our framework embeds a (multi-)neuron point cloud into a fixed-length feature set from which we can decode any point pair affinities, enabling clustering neuron point clouds for automatic proofreading. We also show that the learned feature set can easily be mapped to a contrastive embedding space that enables neuron type classification using a simple KNN classifier. Our approach excels in two demanding connectomics tasks: proofreading segmentation errors and classifying neuron types. Evaluated on three benchmark datasets derived from state-of-the-art connectomes, our method outperforms point transformers, graph neural networks, and unsupervised clustering baselines.

Insights into geometric deviations of medical 3d-printing: a phantom study utilizing error propagation analysis

BackgroundThe use of 3D-printing in medicine requires a context-specific quality assurance program to ensure patient safety. The process of medical 3D-printing involves several steps, each of which might be prone to its own set of errors. The segmentation error (SegE), the digital editing error (DEE) and the printing error (PrE) are the most important partial errors. Approaches to evaluate these have not yet been implemented in a joint concept. Consequently, information on the stability of the overall process is often lacking and possible process optimizations are difficult to implement. In this study, SegE, DEE, and PrE are evaluated individually, and error propagation is used to examine the cumulative effect of the partial errors.MethodsThe partial errors were analyzed employing surface deviation analyses. The effects of slice thickness, kernel, threshold, software and printers were investigated. The total error was calculated as the sum of SegE, DEE and PrE.ResultsThe higher the threshold value was chosen, the smaller were the segmentation results. The deviation values varied more when the CT slices were thicker and when the threshold was more distant from a value of around -400 HU. Bone kernel-based segmentations were prone to artifact formation. The relative reduction in STL file size [as a proy for model complexity] was greater for higher levels of smoothing and thinner slice thickness of the DICOM datasets. The slice thickness had a minor effect on the surface deviation caused by smoothing, but it was affected by the level of smoothing. The PrE was mainly influenced by the adhesion of the printed part to the build plate. Based on the experiments, the total error was calculated for an optimal and a worst-case parameter configuration. Deviations of 0.0093 mm ± 0.2265 mm and 0.3494 mm ± 0.8001 mm were calculated for the total error.ConclusionsVarious parameters affecting geometric deviations in medical 3D-printing were analyzed. Especially, soft reconstruction kernels seem to be advantageous for segmentation. The concept of error propagation can contribute to a better understanding of the process specific errors and enable future analytical approaches to calculate the total error based on process parameters.

Comparison Between Conventional and Artificial Intelligence-Assisted Setup for Digital Implant Planning: Accuracy, Time-Efficiency, and User Experience.

To investigate the reliability and time efficiency of the conventional compared to the automatic artificial intelligence (AI) segmentation of the mandibular canal and registration of the CBCT with the model scan data, in relation to clinician's experience. Twenty clinicians, 10 with a moderate and 10 with a high experience in computer-assisted implant planning, were asked to perform a bilateral localization of the mandibular canal, followed by a registration of the intraoral model scan with the CBCT. Subsequently, for each data set and each participant, the same operations were performed utilizing the AI tool. Statistical significance was assessed via a mixed model (using the PROC MIXED statement and the compound symmetry covariance structure). The mean time for the segmentation of the mandibular canals and the registration of the models was 4.75 (2.03)min for the manual and 2.03 (0.36) min for the AI-automated operations (p < 0.001). The mean discrepancy in the mandibular canals was 0.71 (1.80) mm RMS error for the manual segmentation and 0.68 (0.36) RMS error for the AI-assisted segmentation (p > 0.05). For the registration between the CBCT and the intraoral scans, the mean discrepancy was 0.45 (0.16) mm for the manual and 0.37 (0.07) mm for the AI-assisted superimposition (p > 0.05). AI-automated implant planning tools are feasible options that can lead to a similar or better accuracy compared to the conventional manual workflow, providing improved time efficiency for both experienced and less experienced users. Further research including a variety of software and data sets is required to be able to generalize the outcomes of the present study.

Multivariate Brain Tumor Detection in 3D-MRI Images Using Optimised Segmentation and Unified Classification Model.

3D Magnetic Resonance Imaging (3D-MRI) analysis of brain tumours is an important tool for gathering information needed for diagnosis and disease therapy planning. However, during the brain tumor segmentation process existing techniques have segmentation error while identifying tumor location and extended tumor regions due to improper extraction of initial contour points and overlapping tissue intensity distributions. Hence a novel Duo-step optimised Pyramidal SegNet has been proposed in which multiscale contrast convolutional attention module improve contrast and the tumor edge has been extracted based on location and tumor extension using Duo-step darning needle optimisation that set initial contour points and pyramidal level set segmentation with ancillary Sobel edge operator extract the tumour region from all 2D MRI image slices without having overlapped tissue intensity distributions thereby effectively minimises segmentation error. Furthermore, during the classification of segmented tumor region based on its type, irregular planimetric volume and low interrater concordance of multivariate brain tumors reduce the detection rate due to neglecting the extraction of contextual and symmetric features. Hence 3D brain Unified NN has been proposed in which adaptive multi-layer deep unified encoder module extract 3D contextual and symmetric features by measuring the difference from the observed region and contralateral region and the multivariate brain tumors are classified with boosted Sparse Categorical Cross entropy loss calculation to demonstrate high detection rate. The results obtained for the BraTS2020 and Brain Tumor Detection 2020 data sets showed that the proposed model outperforms existing techniques with excellent precision of 97%, 97.5%, recall of 99%, 97.8%, and accuracy of 95.7%, 98.4%, respectively.

External Validation of Fully-Automated Infrarenal Maximum Aortic Aneurysm Diameter Measurements in Computed Tomography Angiography Scans Using Artificial Intelligence (PRAEVAorta 2).

This study investigates the accuracy of fully-automated maximum aortic diameter measurements in abdominal aortic aneurysm (AAA) patients using artificial intelligence software (PRAEVAorta 2, Nurea, Bordeaux, France). This is a multicenter, retrospective validation study using prospectively collected data from the Zenith alpha for aneurysm Repair Registry (ZEPHYR). Automated measurements of PRAEVAorta 2 are compared with measurements of an internationally recognized core laboratory (Syntactx, New York, New York State). The reviewers at the core laboratory were measurement technologists trained to and utilizing established measurement standards, overseen by vascular surgeons and radiologists. The data set comprised 871 computed tomography angiography scans from the ZEPHYR registry with 347 patients who underwent endovascular aneurysm repair (EVAR) with the Zenith Alpha Endovascular Abdominal Graft (Cook Medical, Bloomington, Indiana) in Germany, Belgium, and The Netherlands between 2016 and 2019. The analysis demonstrated excellent correlation of the measurements (r=0.97) with an intraclass correlation (ICC) of 0.972 (95% confidence interval [CI]=0.968-0.976) across all scans. For preoperative computed tomography (CT) scans, ICC was 0.953 (95% CI=0.941-0.963), and for postoperative scans, ICC was 0.979 (95% CI=0.975-0,983), respectively. In total, 95.4% of measurements were within the clinically acceptable range of 5 mm in absolute difference. In total, 10% of scans demonstrated obvious segmentation errors, mainly due to failure in detecting vessel segments (renal arteries, aortic bifurcation) or due to mis-detecting the outer border of the AAA (duodenum, inferior vena cava, aortic branches) and were excluded from the analysis. In this study, the maximum AAA diameter could be accurately measured fully-automatically by PRAEVAorta 2 (Nurea) in most cases demonstrating that artificial intelligence (AI) software could serve as an important adjunct for research and clinical practice. However, critical review of the generated reports by an experienced observer and cautious use is warranted to identify flawed segmentations. This multicenter, retrospective validation study assessed the accuracy of fully-automated maximum infrarenal aortic aneurysm diameter measurements. It was demonstrated, that 95.4% of measurements were within the clinically acceptable range of 5 mm in absolute difference, positioning the software as a potential adjunct for clinical practice and research. It is also highlighted however, that critical review of the measurements is obligatory, due to a 10% segmentation error rate.

DEUFormer: High‐precision semantic segmentation for urban remote sensing images

AbstractUrban remote sensing image semantic segmentation has a wide range of applications, such as urban planning, resource exploration, intelligent transportation, and other scenarios. Although UNetFormer performs well by introducing the self‐attention mechanism of Transformer, it still faces challenges arising from relatively low segmentation accuracy and significant edge segmentation errors. To this end, this paper proposes DEUFormer by employing a special weighted sum method to fuse the features of the encoder and the decoder, thus capturing both local details and global context information. Moreover, an Enhanced Feature Refinement Head is designed to finely re‐weight features on the channel dimension and narrow the semantic gap between shallow and deep features, thereby enhancing multi‐scale feature extraction. Additionally, an Edge‐Guided Context Module is introduced to enhance edge areas through effective edge detection, which can improve edge information extraction. Experimental results show that DEUFormer achieves an average Mean Intersection over Union (mIoU) of 53.8% on the LoveDA dataset and 69.1% on the UAVid dataset. Notably, the mIoU of buildings in the LoveDA dataset is 5.0% higher than that of UNetFormer. The proposed model outperforms methods such as UNetFormer on multiple datasets, which demonstrates its effectiveness.

Deep learning for 3D vascular segmentation in hierarchical phase contrast tomography: a case study on kidney

Automated blood vessel segmentation is critical for biomedical image analysis, as vessel morphology changes are associated with numerous pathologies. Still, precise segmentation is difficult due to the complexity of vascular structures, anatomical variations across patients, the scarcity of annotated public datasets, and the quality of images. Our goal is to provide a foundation on the topic and identify a robust baseline model for application to vascular segmentation using a new imaging modality, Hierarchical Phase-Contrast Tomography (HiP-CT). We begin with an extensive review of current machine-learning approaches for vascular segmentation across various organs. Our work introduces a meticulously curated training dataset, verified by double annotators, consisting of vascular data from three kidneys imaged using HiP-CT as part of the Human Organ Atlas Project. HiP-CT pioneered at the European Synchrotron Radiation Facility in 2020, revolutionizes 3D organ imaging by offering a resolution of around 20 μm/voxel and enabling highly detailed localised zooms up to 1–2 μm/voxel without physical sectioning. We leverage the nnU-Net framework to evaluate model performance on this high-resolution dataset, using both known and novel samples, and implementing metrics tailored for vascular structures. Our comprehensive review and empirical analysis on HiP-CT data sets a new standard for evaluating machine learning models in high-resolution organ imaging. Our three experiments yielded Dice similarity coefficient (DSC) scores of 0.9523, 0.9410, and 0.8585, respectively. Nevertheless, DSC primarily assesses voxel-to-voxel concordance, overlooking several crucial characteristics of the vessels and should not be the sole metric for deciding the performance of vascular segmentation. Our results show that while segmentations yielded reasonably high scores-such as centerline DSC ranging from 0.82 to 0.88, certain errors persisted. Specifically, large vessels that collapsed due to the lack of hydrostatic pressure (HiP-CT is an ex vivo technique) were segmented poorly. Moreover, decreased connectivity in finer vessels and higher segmentation errors at vessel boundaries were observed. Such errors, particularly in significant vessels, obstruct the understanding of the structures by interrupting vascular tree connectivity. Our study establishes the benchmark across various evaluation metrics, for vascular segmentation of HiP-CT imaging data, an imaging technology that has the potential to substantively shift our understanding of human vascular networks.

Linguistic and Non-linguistic Factors Affecting International Teaching Assistants’ Comprehensibility and Teaching Quality

ABSTRACT Previous studies have examined linguistic and non-linguistic factors affecting the comprehensibility of L2 English speakers. However, there is a relative scarcity of empirical studies specifically investigating how these factors influence the comprehensibility of international teaching assistants (ITAs) as perceived by their undergraduate (UG) students. The current study aimed to fill this gap by examining how linguistic and non-linguistic factors affect American UGs’ judgments of ITAs’ speech comprehensibility and teaching quality. One hundred and ninety-eight U.S. UGs reported their previous experience with and attitudes towards ITAs through a Qualtrics survey, and gave their ratings about ITAs’ speech comprehensibility, segmental and suprasegmental errors, and teaching quality. Results from step-wise regression analyses indicated that segmental errors were statistically significant predictors of UGs’ ratings of ITAs’ comprehensibility and teaching quality, and that UGs who had taken more courses with ITAs had higher ratings of ITAs’ comprehensibility. More notably, UGs’ attitudes towards ITAs were found to contribute to their perceptions of ITAs’ speech comprehensibility and teaching quality. This study concluded with recommendations about how to counter the potential negative effects of biased UG ratings of ITAs so as to promote a healthier environment for ITAs’ professional identity development in higher educational settings.