A two-level AI framework is developed to identify the presence and extent of visible corrosion and classify corrosion pixels into four visual categories. This framework provides complementary information about where corrosion occurs and how accurately its boundaries are represented.
Researchers developed a photospike-based TRNG that harnesses unpredictable light-induced electrical charges to generate true random numbers. The device passed all 15 randomness tests and remained stable over millions of cycles, making it suitable for image authentication and deepfake detection.
A new model called DAC enables medical image segmentation with limited labelled data, achieving consistent generalization across unseen domains. The approach uses feature-level supervision and asymmetric co-training to reduce errors, especially in low-contrast structures.
Researchers developed DeMemSeg, an AI-driven pipeline that accurately segments overlapping membrane structures with accuracy comparable to expert manual analysis. The approach enables large-scale, objective, and quantitative analysis of morphological data, providing a foundational technology for advancing disease mechanisms.
A new AI tool can learn to read medical images with far less data, cutting down the amount of required data by up to 20 times. The tool improves upon medical image segmentation, a labor-intensive task often performed by experts, and boosts model performance in settings with limited annotated data.
An international research team developed a user-friendly software method called Segment Anything for Microscopy, which can precisely segment images of tissues, cells, and similar structures. The new model improved performance for cell segmentation, enabling researchers to automate tasks that previously took weeks of manual effort.
Researchers developed a robust AI model that automates segmenting of MRI images, reducing radiologist workload and improving consistency. The TotalSegmentator MRI model achieved high performance on various anatomical structures, with a Dice score of 0.839.
The new Cellpose3 tool enables easy recognition of cell boundaries in distorted images, restoring crisp images for accurate segmentation. This improvement allows users to segment individual cells with ease, even in conditions with noise, blurring, or undersampling.
A new method called Clio allows robots to make task-relevant decisions by identifying the parts of a scene that matter. In real experiments, Clio successfully mapped scenes at different levels of granularity based on natural-language prompts and enabled robots to grasp objects of interest.
Researchers explore the feasibility of deep learning models in segmenting lesions on PET/CT images to improve salvage radiation therapy planning for prostate cancer. The study demonstrates promising potential to reduce inter- and intra-observer variations, leading to more accurate treatment outcomes.
A new machine learning approach combines computer vision with deep-learning algorithms to pinpoint problem areas in concrete structures. The system enables efficient identification and inspection of cracks using autonomous robots, reducing the overall inspection workload.
CDMENet outperforms fully and semi-supervised counterparts in grape yield prediction using Mean Absolute Error (MAE) and Root Mean Square Error (RMSE). The algorithm's robustness is demonstrated with limited labeled data, highlighting its potential as a cost-effective tool in agricultural yield estimation.
Novel Dice loss functions, t-vMF Dice loss and Adaptive t-vMF Dice loss, have been developed to improve image segmentation accuracy in medical images. These new functions outperform conventional formulations and show great potential for critical fields like medical imaging and diagnosis.
Researchers at Osaka University develop a method to train AI models using simulated city images, reducing the need for real data and saving human effort. The approach generates realistic images with accurate ground truth labels, addressing instance segmentation challenges.
DragGAN enables non-professionals to perform complex image edits with AI support, adjusting pose, gaze direction, and viewing angle. The method uses Generative Adversarial Networks to generate new images, promising simplified post-processing for AI-generated content.
Researchers at North Carolina State University have developed a new methodology called Patch-to-Cluster attention (PaCa) that addresses the challenges of vision transformers. PaCa improves ViT's ability to identify, classify, and segment objects in images while reducing computational demands and enhancing model interpretability.
A new technique called image-free single-pixel object detection (SPOD) can detect the location, size, and category of multiple objects without acquiring images. SPOD uses a small optimized structured light pattern to quickly scan the scene and extract features, achieving an accuracy of over 80%.
A novel AI architecture, relational reasoning network, accurately identifies anatomical landmarks in CT scans for orthodontic treatments. The model learns spatial relationships between landmarks without explicit image segmentation, achieving accuracy comparable to conventional methods.
The new automated method, called HybridFlow, uses large-scale aerial images to produce precise 3D models of cityscapes and landscapes. This technology has potential applications in natural disaster risk assessment and mitigation, enabling informed decision-making and evaluation of risk-mitigating factors.
Researchers have developed a diffractive optical processor that can compute hundreds of transformations in parallel using wavelength multiplexing. The processor, which is powered by light instead of electricity, can execute multiple complex functions simultaneously at the speed of light.
A team of researchers from Korea investigated the dynamics of the p-Laplacian AC equation, finding that solutions maintain three criteria: phase separation, boundedness, and energy decay properties. They also identified an advantage of p-AC equation over classical Laplacian in adjusting interface sharpness.
City digital twin technology is used to create synthetic training data for deep learning models, which are then trained on a combination of real and synthetic data. This approach yields promising results for architectural segmentation tasks, particularly for modern building styles.
A team of researchers at Osaka University has created a machine learning system that can virtually remove buildings from a live view, streaming in real-time on a mobile device. This technology can help accelerate the process of urban renewal based on community agreement, reducing conflicts and delays.
A newly expanded data set of brain scans from stroke patients called ATLAS now includes 1,271 MRI images with manually segmented lesions, facilitating large-scale stroke recovery research. Researchers hope to develop algorithms to automate lesion segmentation, enabling clinicians to predict patient responses to therapies.
A team from KAUST has developed a low-cost system for imaging plant growth dynamics noninvasively and at high throughput. The Mutiple XL ab system combines computer vision and pattern recognition technologies with machine learning to analyze and quantify root growth dynamics.
Researchers have developed a new training method for machine learning models to perform blood cell counts, reducing manual annotation work. The U-Net model achieves high accuracy in segmenting images with multiple cell types, promising a simpler and cheaper alternative to traditional cell analyzers.
A prospective study found that the 3-T Dixon gradient-recalled echo (GRE) sequence performed better than the current standard of 1.5-T SSFP for unenhanced coronary MRA, particularly in distal and branch segments. The technique demonstrated higher image quality, visible segments, sensitivity, and specificity for significant stenoses.
A recent study uses deep neural networks to analyze CT scans of dinosaur fossils, reducing manual segmentation time from days to minutes.
A new AI-based method has been developed to evaluate patients with Stargardt disease, a leading cause of childhood blindness. The study found that the severity of vision loss can be classified into different phenotypes based on genetic variants, and provided sensitive structural outcome measures for therapeutic trials.
Automated brain volumetry in memory-impaired patients shows significant differences and systematic biases between conventional and ultrafast 3D T1-weighted MRI sequences. Most regions demonstrated substantial agreement but also significantly different mean values and consistent biases.
A novel 'virtual segmentation' method enables accurate visualization of microtomography imaging of Egyptian mummies. This technique helps researchers reconstruct detailed anatomical structures of ancient animals, shedding new light on their biology and evolution.
A new machine learning algorithm has enabled researchers to automatically identify and map the inner structures of cells, including organelles, with unprecedented precision. By processing tens of thousands of high-resolution images, scientists have gained insights into how these structures interact and are arranged within the cell.
Researchers developed a novel approach to 3D image segmentation, segmenting the gaps between parts instead of contours, to automate tedious tasks. The technique demonstrates promising results in diagnosing TMJ-related issues and has potential applications in other fields.
The new EQUIPS workflow provides a more accurate and reliable way to process 3D images for computer simulations. It uses machine learning to automate the drawing process and produces a range of simulation outcomes, allowing decision-makers to consider best- and worst-case scenarios.
A team of researchers at Osaka University created a custom dataset to train an AI algorithm to digitally remove unwanted objects from building façade images. The algorithm achieved high accuracy in inpainting occluded regions with digital inpainting.
A team of scientists from Osaka University developed a machine learning method for classifying the type of building and its primary façade color using deep learning models applied to street-level images. This work may assist in fostering neighborhood cohesion and support urban renewal by providing tailored street-view datasets.
The team used machine learning technique generative adversarial networks to digitally remove clouds from aerial images, generating accurate datasets of building image masks. This work may help automate computer vision jobs critical to civil engineering, enabling the detection of buildings in areas without labeled training data.
A novel computational tool called CShaper speeds up the analyzing process from hundreds of hours to a few hours by computer. It can segment and analyze cell images systematically at the single-cell level, enabling biologists to decipher the contents of these images within a few hours.
A new AI technique detects acute ischemic stroke lesions on MRIs with high accuracy, outperforming expert drawn gold standard. The fully automated approach reduces workflow time and operator bias in lesion segmentation.
A new AI tool uses deep learning to automate the segmentation of individual muscles from CT images, enabling the creation of personalized musculoskeletal models. This advancement has significant implications for patients with musculoskeletal diseases, such as ALS, and high-performance athletes seeking to improve their performance.
A new system uses a single labeled scan and unlabeled scans to automatically generate a large dataset of distinct training examples. The system's approach learns anatomical, brightness, and contrast variations from unlabeled scans to synthesize realistic and accurately labeled scans.
The MIT-developed AI model can associate specific words with specific patches of pixels in an image, enabling real-time object highlighting based on spoken descriptions. This innovation holds promise for applications such as language translation and automatic image annotation.
Researchers at MIT's CSAIL have developed RoadTracer, an automated method to build road maps that is 45 percent more accurate than existing approaches. The system uses data from aerial images and creates maps step-by-step, tracing out roads one step at a time.
Researchers at North Carolina State University developed a new image segmentation technique that improves object identification and separation in images. The technique, called Consensus-Based Image Segmentation via Topological Persistence, aggregates data from multiple algorithms to create a new version of the image.
A novel wavelet variational model is proposed to segment ultrasound videos efficiently, tackling low contrast, shadow effects, and complex noise statistics. The model achieves accurate ROI tracking with robustness and flexibility, making it suitable for real-time clinical applications.
A team led by Berkeley Lab scientist Gang Ren captured the first 3-D images of individual double-helix DNA segments attached to gold nanoparticles. The images reveal the flexible structure of the DNA segments, which could aid in building molecular devices for drug delivery, biological research, and electronic devices.
A new algorithm developed by MIT researchers combines SLAM and object recognition to improve robots' performance. The system uses SLAM information to augment existing object-recognition algorithms, achieving comparable performance to special-purpose robotic object-recognition systems that factor in depth measurements.
Researchers at Disney Research Zurich developed a method for accurate video object segmentation using a click-and-drag interface that enables human editors to work efficiently with state-of-the-art algorithms. This approach achieves higher accuracies than fully automated systems by keeping a human in the loop.
A new image analysis and segmentation algorithm helps quantify grain sizes in Martian rock and soil images, revealing subtle trends with composition. The semi-automated algorithm provides better consistency across multiple images than human segmentation.
Researchers have developed new superresolution and segmentation methods for magnetic resonance images to analyze structural brain differences in psychotic patients and their healthy relatives. These methods improve the quality of images and enable automatic calculations of desired sizes, leading to a better understanding of psychosis.
Engineers are using COCOA to verify the correct shape of the mirrors on NASA's James Webb Space Telescope, which will be tested at operating temperatures. The results will ensure the telescope works correctly as a whole, providing images of the first galaxies and exploring planets around distant stars.
A new cardiac CT technique, prospective gated 64-channel cardiac CT, has been shown to significantly reduce patient exposure to radiation. Compared to the standard retrospective ECG gating technique, this method produces high-quality images with a 76% lower radiation dose.
Researchers use Google Sets to provide contextual information that improves the accuracy of automated image labeling systems. The system uses a three-step process, including image segmentation, ranked lists of probable labels, and post-processing context checks.
Researchers at Carnegie Mellon University have developed two systems that use web images to enhance edited photos. Photo Clip Art uses labeled images from LabelMe as clip art, while Scene Completion draws upon millions of photos from Flickr to fill in holes. These systems enable users to achieve realistic results with minimal skills.