Annotation-efficient segmentation of the numerous mitochondria instances from various electron microscopy (EM) images is highly valuable for biological and neuroscience research. Although unsupervised domain adaptation (UDA) methods can help mitigate domain shifts and reduce the high costs of annotating each domain, they typically have relatively low performance in practical applications. Thus, we investigate weakly supervised domain adaptation (WDA) that utilizes additional sparse point labels on the target domain, which require minimal annotation effort and minimal expert knowledge. To take full use of the incomplete and imprecise point annotations, we introduce a multitask learning framework that jointly conducts segmentation and center detection with a novel cross-teaching mechanism and class-focused cross-domain contrastive learning. While leveraging unlabeled image regions is essential, we introduce segmentation self-training with a novel instance-aware pseudo-label (IPL) selection strategy. Unlike existing methods that typically rely on pixel-wise pseudo-label filtering, the IPL semantically selects reliable and diverse pseudo-labels with the help of the detection task. Comprehensive validations and comparisons on challenging datasets demonstrate that our method outperforms existing UDA and WDA methods, significantly narrowing the performance gap with the supervised upper bound. Furthermore, under the UDA setting, our method also achieves substantial improvements over other UDA techniques.

Cryo-electron tomography (cryoET) enables visualization of macromolecular complexes within intact cellular environments. Continued improvements in instrumentation, sample preparation, and data-processing pipelines have increased both the scale and the complexity of cryoET datasets, making manual analysis challenging. To support scalable, collaborative annotation, we developed copick, an open-source dataset application programming interface (API) and accompanying tool suite for cryoET analysis. Copick provides standardized access to tomograms, segmentations, point annotations, meshes, and feature maps across local storage, high-performance computing systems, cloud platforms, and public repositories. Plugins for napari and ChimeraX enable human-in-the-loop workflows for particle picking, segmentation, inspection of machine-learning outputs, and project-level collaboration. A multi-resolution Open Microscopy Environment (OME)-Zarr architecture supports responsive visualization and cross-platform access. Copick additionally provides a Model Context Protocol interface enabling automated generation of annotation-curation pipelines using natural-language instructions. Together, these tools support reproducible, scalable, and collaborative cryoET analysis.

Background/Objectives: Precise nuclei instance segmentation is a prerequisite for reliable digital pathology, yet the scarcity of pixel-level annotations remains a significant bottleneck for deep learning models. Methods: We propose a self-evolving framework for robust nuclei segmentation that uses only sparse point annotations, extending the Segment Anything Model (SAM). To overcome the limitations of static pseudo-labels, our method introduces a self-evolving labeling strategy via Exponential Moving Average (EMA), which adaptively refines learning targets. We also integrate instance-aware contrastive learning using point prompts as spatial anchors and implement a consensus-based filtering mechanism between prompt-guided and prompt-free decoders. Results: Extensive evaluations on CPM17, MoNuSeg, and the challenging CoNSeP datasets demonstrate that our framework achieves state-of-the-art performance across various backbones, including ViT-B and ViT-H. Conclusions: By enabling a seamless transition from general-purpose foundation models to specialized histopathology experts, this self-refining approach delivers a highly efficient, accurate solution for automated diagnostic workflows in clinical settings.

According to the complexity and diversity of AI aesthetic education course resources and other characteristics, this paper designs an intelligent management and recommendation system for AI aesthetic education course resources, which realizes the intelligent management of AI aesthetic education course resources through the knowledge point annotation algorithm based on TextCNN-Transformer, and designs a multi-factor fusion collaborative filtering recommendation algorithm after the resource knowledge point annotation task. . The recommendation algorithm of multi-factor fusion collaborative filtering implements the intelligent recommendation of AI aesthetic education course resources by constructing the AI aesthetic education course resources knowledge graph, user knowledge graph, using the interest similarity calculation to obtain the recommendation score ranking of the AI aesthetic education course resources matched with the user's interests, and using the recommendation algorithm of multi-factor fusion collaborative filtering to implement the intelligent recommendation of AI aesthetic education course resources. The evaluation index system of the intelligent management and recommendation system of AI aesthetic education course resources is constructed, and the fuzzy evaluation method is combined to obtain the final score of the intelligent management and recommendation system of AI aesthetic education course resources designed in this paper.The performances of the TextCNN-Transformer model in the HM Loss, Sub Acc, Macro F1, and Micro F1 indexes were 0.0125 The RMSE and MAE of the recommendation algorithm of multi-factor fusion collaborative filtering on Filmtrust dataset reached 0.651 and 0.441 respectively.The fuzzy evaluation score of AI aesthetic education curriculum resources intelligent management and recommendation system is 84.646, which is in the interval of [80,90], with a good operational efficiency and a good construction of intelligent management and recommendation system is well constructed.

Mitochondrial morphology is a critical indicator of cellular metabolic status and disease pathogenesis, requiring high-resolution visualization and precise segmentation in electron microscopy (EM) images. While fully supervised deep learning models have achieved significant progress, their reliance on dense pixel-wise annotations presents a major bottleneck due to the labor-intensive labeling process and expert variability near the optical diffraction limit. Existing weakly supervised methods, primarily designed for densely packed instances, often fail to generalize to the sparse distribution of mitochondria in EM data. In this paper, we propose WeakMitoSAM, a novel weakly supervised framework for high-precision mitochondria segmentation using sparse point annotations. Our approach introduces the competitive aggregation of multiple prompts strategy, which employs a Bias-augmented Softmax mechanism to reconcile semantic ambiguities and suppress background noise, effectively converting sparse priors into high-fidelity pseudo-labels. Subsequently, segment anything model is specialized for mitochondrial ultrastructures via low-rank adaptation, ensuring parameter-efficient domain adaptation. Experimental results across four public EM datasets demonstrate that WeakMitoSAM achieves state-of-the-art performance in point-supervised scenarios and even outperforms several fully supervised benchmarks, providing an efficient and robust solution for large-scale mitochondrial morphofunctional analysis.

Facial nerve damage remains a significant risk during vestibular schwannoma (VS) resection, with reported incidences varying widely (3-46%). Damage risk increases with tumor size. Digital tractography enables nerve reconstruction but typically involves manual procedures, resulting in subjective evaluations that limit reproducibility and validation. We introduce a robust, semi-automatic tractography methodology with reproducible region-of-interest (ROI) generation and present an initial validation using a novel quantitative three-dimensional comparison approach in patients with a large VS. To assess the accuracy of facial nerve reconstruction employing a semi-automatic ROI selection method in patients with VSs. We included six patients with an average tumor size of 28 mm (95% CI: 17-40, 100% left) who underwent translabyrinthine VS surgery. Each VS patient was scanned with the regular neuronavigation magnetic resonance imaging (MRI) protocol and a custom diffusion-MRI protocol before surgery. The facial nerve trajectory was reconstructed with a diffusion tensor imaging-based tractography software package using semi-automatic ROI generation. We validated our reconstructions with the Brainlab neuronavigation system for intraoperative point annotation along the course of the facial nerve. Tracts could be reconstructed in all included patients. The median distance and angle between the points and closest reconstruction were 5.1 mm (IQR: 3.5-7.6) and 38.5 degrees (IQR: 2.7-79.8), respectively. We present a promising methodology for facial nerve reconstruction in patients with VSs. However, further optimization of the methodology is warranted before a proper clinical validation study can be performed.

This study addresses the challenge of automated high-throughput phenotyping of wheat spike characteristics using modern computer vision and deep learning methods. Accurate estimation of spikelet number is a key indicator of plant productivity, yet traditional manual counting approaches are labor-intensive, slow, and difficult to scale to large breeding datasets. To overcome these limitations, we propose a spikelet detection strategy based on simplified point annotations, where an expert marks only the centers of spikelets rather than drawing detailed segmentation masks or bounding boxes. This significantly reduces annotation time and lowers the overall cost of preparing training datasets for machine learning models. To determine the most effective way of utilizing such simplified annotations, three computational methods were explored: segmentation of binary masks using a U-Net architecture, density regression based on two-dimensional Gaussian distributions optimized via Kullback-Leibler divergence, and detection of fixed-size bounding regions using the YOLOv8 object detection framework. The models were evaluated on dedicated test datasets using both quantitative metrics (MAE, MAPE) and spatial localization metrics (Precision, Recall, F1 score). The results demonstrate that U-Net-based approaches provide consistently high accuracy in spikelet localization and counting while maintaining robustness to annotation imperfections. In contrast, the YOLOv8-based method showed reduced performance, likely due to the geometric mismatch between fixed-size boxes and the natural elongated shape of spikelets. Overall, the proposed methodology highlights the effectiveness of combining minimalistic point-level annotation with advanced segmentation models for automating phenotyping workflows. This approach has the potential to accelerate breeding programs, enhance the efficiency of large-scale phenotypic data collection, and support further development of robust computer-vision tools for plant science applications.

"Facial Beauty" is not an absolute physical attribute but a subjective social and cultural construct. Facial beauty assessment is an interdisciplinary field that integrates computer vision and medical aesthetics (MAs) to quantify personal judgment regarding facial attractiveness. In this study, the beauty assessment we adopted was based on the scores given by plastic surgeons; this method is more professional and is supported by a theoretical basis. We derived a set of MA features that encompass global traits, local details, and curvature aspects from established aesthetic principles. Incorporating these features enhances predictive accuracy in facial beauty. Furthermore, we propose a feature selection algorithm with aesthetic-driven initialization embedded in a multi-objective evolutionary framework. Additionally, we introduce an MA facial landmark model that provides explicit annotation of bilateral zygomatic, orbital, and nasal points for precise attractiveness scoring. Experimental results on the South China University of Technology-Facial Beauty Perception (SCUT-FBP) and SCUT-FBP5500 datasets and the Chicago Face Dataset demonstrate superior performance (Pearson's correlation coefficient = 0.8216, mean absolute error = 0.2638, and root mean square error = 0.3743) over state-of-the-art methods, validating its clinical relevance. This study provides a practical tool for beauty evaluation, where the selected features align with professional judgments, enabling transparent and explainable outcomes in both clinical and cosmetic applications.

Retinal vein occlusion (RVO) is one of the most common vision-threatening retinal diseases, with macular edema (ME) as its primary complication. Optical coherence tomography (OCT), a non-invasive imaging modality, enables detailed visualization of retinal structures and fluid distribution, thus supporting accurate diagnosis, treatment monitoring, and clinical assessment of RVO-related conditions. However, the development of automated algorithms for RVO-ME analysis has been hindered by the lack of high-quality, manually segmented datasets. To address this limitation, we constructed a manually annotated RVO-ME dataset comprising 3,012 OCT B-scans from 146 eyes of 130 patients. For each image, we provide segmentation labels for four key retinal features (subretinal fluid, intraretinal fluid, the ellipsoid zone, and the external limiting membrane), along with point annotations to facilitate the detection of highly reflective foci. This dataset provides a valuable benchmark for assessing the performance of segmentation algorithms and facilitates the advancement of artificial intelligence models for RVO-related disease analysis.

Benthic ecological surveys yield a massive volume of seabed imagery, yet analyzing the abundance of organisms remains a time-consuming task for experts. This bottleneck hinders the analysis of all collected data. Convolutional Neural Networks (CNNs) offer a promising solution for automating image analysis. However, training CNNs requires images with bounding boxes drawn around the target organisms. Such datasets are often unavailable, as prior research primarily relied on manual point annotations for organism locations. This study presents a novel workflow for training CNN to identify benthic organisms using existing point annotations. We demonstrate that legacy point annotations from previous surveys can be used to annotate new images collected within the same study area. Our results show that the CNN's predictions were comparable to discrepancies found in inter-expert variability. While the accuracy may not surpass models trained with dedicated bounding box datasets, our approach proves that historical point annotations can effectively generate training data for object detection CNNs, particularly when dedicated bounding box datasets are scarce. Given the vast number of past and ongoing benthic surveys utilizing point annotations, this approach unlocks new avenues for machine learning in marine ecology. • Legacy point annotations can be reused to train CNNs for deep-sea image analysis. • An open-source workflow enables non-experts to replicate the training pipeline. • Accuracy of CNN predictions is comparable to the variability seen between human expert annotations. • This approach reduces manual effort and speeds up deep-sea image data processing. • Case Study demonstrates potential of using existing annotated datasets for scalable marine AI research.

Light Field Salient Object Detection (LFSOD) aims to identify visually distinctive regions by leveraging the complementary spatial-angular information inherent in 4D light field imagery. A major challenge lies in modeling angular dependencies and maintaining spatial coherence under sparse supervision. In this paper, we propose a weakly supervised network that consists of three interdependent modules. First, the Light Field Division (LFD) module utilizes epipolar geometry to extract direction-aware boundary features, enhancing the encoding of angular disparities. Second, the Light Field Spatial Association (LFSA) module anchors cross-view feature alignment using central-view point annotations, thereby enforcing spatial consistency and mitigating redundant representations. Third, the Light Field Saliency Local Clustering (LFLC) module introduces a joint boundary-appearance modeling strategy that integrates adaptive clustering with error-aware regularization to refine structural predictions. Experiments on three benchmark datasets show that our method consistently outperforms mainstream weakly supervised approaches. It also achieves superior performance compared to several fully supervised methods.

Temporal Sentence Grounding (TSG) in videos aims to localize a temporal interval from an untrimmed video that is semantically relevant to a given query sentence. To achieve a balance between tremendous annotation burden and grounding performance, we propose a new Weakly Semi-supervised Temporal Sentence Grounding with Points (WSS-TSG-P) task, where the dataset comprises limited fully-annotated video-sentence pairs by start and end timestamps (full label) and a large amount of weakly-annotated pairs by a single point timestamp (point label). Based on this setting, we first introduce a point-tomoment1 regressor which converts point annotations to pseudo moment labels. To train a good regressor for reliable pseudo moment labels, we propose a point-guided feature aggregation module to aggregate cross-modal representations based on the prototype feature at the given point position. In addition, we propose to perform regressor self-training and design pseudo label generation strategies to exploit both full annotations and point annotations. All heterogeneous labels (full, pseudo moment, and point labels) are used to train a TSG backbone. In addition, we propose a novel point-guided group contrastive learning method by constructing reliable positive and negative sets and re-weighting pseudo moment labels to further improve the model performance. Extensive experiments on benchmark datasets verify that our proposed method outperforms other semi-supervised learning methods and bridges the performance gap between weakly-supervised and fully-supervised learning methods in TSG.

ABSTRACT The pitch and row pitch of rock bolts are critical parameters in rock‐support techniques. These parameters facilitate tracking construction processes and monitoring bolt displacement for early warning of rock bursts and roof falls. A measurement method for these pitches based on dual‐modal image recognition is proposed to solve the problems of high personnel safety risks, low accuracy, and high labour intensity in traditional measurement. First, an intrinsically safe structured‐light camera captures colour images and point clouds, and Gaussian heatmap‐based point annotation is conducted on the colour images to produce training samples. Second, the detection network is enhanced via point feature maps, feature point attention, and mixed up‐sampling. A weighted Gaussian heatmap is adopted as the loss function to develop a rock bolt point detection network, which effectively enhances the detection and localisation accuracy. Finally, a two‐stage serial fitting method is proposed to solve the linear equation for each row and complete the measurement. Simulation experiments demonstrate that the developed point detection network achieves notable gains in precision, recall, and positioning error. The precision is 87.53%, the recall is 93.03%, and the positioning error is 11.33 mm, all ranking among the top performers. Industrial experiments demonstrate that the method is well‐suited for on‐site applications.

Deep convolutional neural networks have demonstrated remarkable effectiveness in image segmentation. However, segmentation becomes challenging when training on images with complex instances. Moreover, obtaining annotations for high-precision data is also difficult. Weakly supervised learning can address this issue by using nonspecialised annotations or supervised information from segmentation algorithms. In this study, we proposed TSSP-UNet: a two-stage weakly supervised segmentation approach. In the first stage, we trained a segmentation network augmented with constraint and attention mechanisms. These mechanisms are designed to operate on boundaries and superpixels generated from pseudo-labels. For the attention network, two pseudo-labels were used with a binary mask to add contour information to the segmentation process. Furthermore, a feature aggregation segmentation network was applied to the prominent foreground area in the image by incrementally adding elements. In the second stage, a refined confident learning algorithm improved the pseudo-labels at the pixel level and then TSSP-UNet was retrained using the modified superpixel labels. Testing on the MoNuSeg and TNBC datasets demonstrates that the approach performs well in the weakly supervised cell nucleus segmentation task compared with baseline methods.

Single-point annotation is increasingly prominent in visual tasks for labeling cost reduction. However, it challenges tasks requiring high precision, such as the point-prompted instance segmentation (PPIS) task, which aims to estimate precise masks using single-point prompts to train a segmentation network. Due to the constraints of point annotations, granularity ambiguity and boundary uncertainty arise $ {\mathit{i.e.}}$, the difficulty distinguishing between different levels of detail ($ {\mathit{e.g.}}\,$ whole object vs. parts) and the challenge of precisely delineating object boundaries. Previous works have usually inherited the paradigm of mask generation along with proposal selection to achieve PPIS. However, proposal selection relies solely on category information, failing to resolve the ambiguity of different granularity. Furthermore, mask generators offer only finite discrete solutions that often deviate from actual masks, particularly at boundaries. To address these issues, we propose the Semantic-Aware Point-Prompted Instance Segmentation Network (SAPNet). It integrates Point Distance Guidance and Box Mining Strategy to tackle group and local issues caused by the point's granularity ambiguity. Additionally, we incorporate completeness scores within proposals to add spatial granularity awareness, enhancing multiple instance learning (MIL) in proposal selection termed S-MIL. The Multi-level Affinity Refinement conveys pixel and semantic clues, narrowing boundary uncertainty during mask refinement. These modules culminate in SAPNet++, mitigating point prompt's granularity ambiguity and boundary uncertainty and significantly improving segmentation performance. Extensive experiments on four challenging datasets validate the effectiveness of our methods, highlighting the potential to advance PPIS.

IntroductionThe application of deep learning methods in automatic delineation of fiducial points in seismocardiography (SCG) on a beat-to-beat basis provides the possibility of obtaining a novel and comprehensive approach to assess and monitor myocardial mechanics and hemodynamic status. Therefore, the aim of this study was to develop an adaptive and data-driven algorithm for automatic delineation of 11 fiducial points in SCG.MethodsSCG signals from subjects both with and without known cardiac disease (CD) were included. A semi-automatic annotation pipeline was prepared for effective annotation of fiducial points for each individual cardiac cycle, in which 42,452 individual beats from 198 subjects were annotated. A deep learning model with U-Net architecture was developed to detect 11 fiducial points and predict multiple time intervals in the SCG signal. The evaluation metrics were positive predictive value and sensitivity.ResultsThe median positive predictive value and sensitivity of the algorithm ranged between 0.809 and 1.000 and 0.843 and 0.918 for different fiducial points, respectively.ConclusionA novel algorithm for automatic detection of 11 fiducial points in SCG was developed and tested in subjects both with and without CD.

Small object localization is one of the most challenging tasks owing to the poor visual appearance and noisy representation caused by the intrinsic structure of small targets. Recent advances in localizing small objects are mainly dependent on regression-based counting approaches, which require considerable annotations for training. As a contrast, human learners can quickly master labeling skills from only a few annotation examples. In this paper, we attempt to simulate this training mechanism and propose a novel positive-unlabeled (PU) learning based approach that can localize small objects by learning from partial point annotations. We evaluate our approach on five typical datasets of small objects involving a single cell, an animal/insect, and human crowds. Quantitative experimental results show that our approach has achieved inspiring localization performance (F1 score > 0.75) even under the supervision of less than 10% of the overall point annotations. This approach paves the way for low-annotation-cost single-cell analysis within microfluidic droplets.

SP-DETR: Superior point weak semi-supervised DETR with teacher–student paradigm for crop and weed detection

Nuclei segmentation is crucial for cancer di-agnosis but faces high annotation costs due to dense nu-clei distribution. Weakly supervised learning with point an-notations alleviates this burden, yet single-center data is limited, and centralized datasets are hindered by privacy concerns. Federated learning enables multi-institution col-laboration while preserving privacy, but non-IID data dis-tribution—particularly style heterogeneity from staining and equipment variations—complicates model aggregation. In this paper, we propose Federated learning with Style Perturbation and Clustering (FedSPC), a novel framework that integrates a Federated Style Perturbation (FedSP) model and a Federated Style Clustering (FedSC) strategy. During training, FedSP applies style adversarial perturbation to extract and adapt local style features, reducing local style bias. Meanwhile, FedSC groups clients by style similarity and adjusts aggre-gation weights based on intra-group performance, mitigating fairness propagation bias. FedSPC overcomes pathological image style heterogeneity through the combined use of FedSP and FedSC, delivering a practical federated learning solution for medical imaging. Evaluated against existing federated weakly supervised frameworks, conventional methods, and aggregation schemes, our approach significantly outperforms alternatives in nuclei segmentation tasks. Experiments con-firm FedSPC’s superiority in handling style diversity and improving segmentation accuracy under federated settings.

Abstract Developing weak annotations is one of the effective methods to reduce the workload of labeling the training data and extract the valuable priori information. In this paper, we propose a novel weakly supervised variational segmentation model that requires only a single randomly selected point within the target region. First, We construct an anisotropic Riemannian metric that incorporates not only the texture patterns, the local directional information at the boundaries, but also the intensity heterogeneity through a conformal factor. This step is the cornerstone of our method, as it enables the effective integration of multi-scale image features. Second, we calculate two types of distances: geodesic distances based on the new Riemannian metric and Euclidean distances, from all the points in the image to the taken point. These pointwise distances are then integrated into the Chan-Vese model, enabling the simultaneous exclusion of regions with similar intensities both near and far from the target region. The distances serve as adaptive weighting parameters in the variational model, further enhancing segmentation accuracy. We employ the Douglas-Rachford algorithm for efficient numerical implementation. Experimental results show that our method achieves similar or better segmentation performance than other weakly annotated methods and the single-point annotation has better robustness.