ArXiv [Preprint]. 2025 May 20:arXiv:2505.14507v1.

ABSTRACT

BACKGROUND: Deep learning has potential to improve the efficiency and consistency of radiation therapy planning, but clinical adoption is hindered by the limited model generalizability due to data scarcity and heterogeneity among institutions. Although aggregating data from different institutions could alleviate this problem, data sharing is a practical challenge due to concerns about patient data privacy and other technical obstacles.

PURPOSE: This work aims to address this dilemma by developing FedKBP+, a comprehensive federated learning (FL) platform for predictive tasks in real-world applications in radiotherapy treatment planning.

METHODS: We implemented a unified communication stack based on Google Remote Procedure Call (gRPC) to support communication between participants whether located on the same workstation or distributed across multiple workstations. In addition to supporting the centralized FL strategies commonly available in existing open-source frameworks, FedKBP+ also provides a fully decentralized FL model where participants directly exchange model weights to each other through Peer-to-Peer communication. We evaluated FedKBP+ on three predictive tasks using scale-attention network (SA-Net) as the predictive model.

CONCLUSIONS: Our results demonstrate that FedKBP+ is highly effective, efficient and robust, showing great potential as a federated learning platform for radiation therapy.

PMID:40470470 | PMC:PMC12136487

Front Plant Sci. 2025 May 21;16:1575038. doi: 10.3389/fpls.2025.1575038. eCollection 2025.

ABSTRACT

The rising costs and labor shortages have sparked interest in machine harvesting of fresh-market blueberries. A major drawback of machine harvesting is the occurrence of internal bruising, as the fruit undergoes multiple mechanical impacts during this process. Evaluating fruit internal bruising manually is a tedious and time-consuming process. In this study, we leveraged deep learning models to rapidly quantify berry fruit internal bruising. Blueberries from 61 cultivars of soft to firm types were subjected to bruise over a three-year period from 2021-2023. Dropped berries were sliced in half along the equator and digitally photographed. The captured images were first analyzed using the YOLO detection model to identify and isolate individual fruits with bounding boxes. Then YOLO segmentation models were performed on each fruit to obtain the fruit cross-section area and the bruising area, respectively. Finally, the bruising ratio was calculated by dividing the predicted bruised area by the predicted cross-sectional area. The mean Average Precision (mAP) of the bruising segmentation model was 0.94. The correlation between the bruising ratio and ground truth was 0.69 with a mean absolute percentage error (MAPE) of 15.87%. Moreover, analysis of bruising ratios of different cultivars revealed significant variability in bruising susceptibility and the mean bruising ratio of 0.22 could be an index to differentiate the bruise-resistant and bruise-susceptible cultivars. Furthermore, the mean bruising ratio was negatively correlated with mechanical texture parameter, Young's modulus 20% Burst Strain. Overall, this study presents an effective and efficient approach with a user-friendly interface to evaluate blueberry internal bruising using deep learning models, which could facilitate the breeding of blueberry genotypes optimized for machine harvesting. The models are available at https://huggingface.co/spaces/c-tan/blueberrybruisingdet.

PMID:40470370 | PMC:PMC12133755 | DOI:10.3389/fpls.2025.1575038

Front Plant Sci. 2025 May 21;16:1588901. doi: 10.3389/fpls.2025.1588901. eCollection 2025.

ABSTRACT

INTRODUCTION: The accurate identification of maize varieties is of great significance to modern agricultural management and breeding programs. However, traditional maize seed classification methods mainly rely on single modal data, which limits the accuracy and robustness of classification. Additionally, existing multimodal methods face high computational complexity, making it difficult to balance accuracy and efficiency.

METHODS: Based on multi-modal data from 11 maize varieties, this paper presents DualCMNet, a novel dual-branch deep learning framework that utilizes a one-dimensional convolutional neural network (1D-CNN) for hyperspectral data processing and a MobileNetV3 network for spatial feature extraction from images. The framework introduces three key improvements: the HShuffleBlock feature transformation module for feature dimension alignment and information interaction; the Channel and Spatial Attention Mechanism (CBAM) to enhance the expression of key features; and a lightweight gated fusion module that dynamically adjusts feature weights through a single gate value. During training, pre-trained 1D-CNN and MobileNetV3 models were used for network initialization with a staged training strategy, first optimizing non-pre-trained layers, then unfreezing pre-trained layers with differentiated learning rates for fine-tuning.

RESULTS: Through 5-fold cross-validation evaluation, the method achieved a classification accuracy of 98.75% on the validation set, significantly outperforming single-modal methods. The total model parameters are only 2.53M, achieving low computational overhead while ensuring high accuracy.

DISCUSSION: This lightweight design enables the model to be deployed in edge computing devices, allowing for real-time identification in the field, thus meeting the practical application requirements in agricultural Internet of Things and smart agriculture scenarios. This study not only provides an accurate and efficient solution for maize seed variety identification but also establishes a universal framework that can be extended to variety classification tasks of other crops.

PMID:40470359 | PMC:PMC12133733 | DOI:10.3389/fpls.2025.1588901

Data Brief. 2025 May 2;60:111610. doi: 10.1016/j.dib.2025.111610. eCollection 2025 Jun.

ABSTRACT

Advancements in computer vision and deep learning have transformed ecological monitoring and species identification, enabling automated and accurate data labelling. Despite these advancements, robust AI-driven solutions for avian species recognition remain limited, primarily due to the scarcity of high-quality annotated datasets. To address this gap, this article introduces IBERBIRDS-a comprehensive and publicly accessible dataset specifically designed to facilitate automatic detection and classification of flying bird species in the Iberian Peninsula under real-world conditions. The dataset comprises 4000 images representing 10 ecologically significant medium to large-sized bird species, with each image annotated using bounding box coordinates in the YOLO detection format. Unlike existing datasets that typically feature close-up or ideal-condition imagery, IBERBIRDS focuses on mid-to-long range photographs of birds in flight, providing a more realistic and challenging representation of scenarios commonly encountered in birdwatching, conservation, and ecological monitoring. Images were sourced from publicly available, expert-validated ornithology platforms and underwent rigorous quality control to ensure annotation accuracy and consistency. This process included homogenizing color profiles and formats, as well as manual refinement to ensure that each image contains a single bird specimen. Additionally, detailed provenance and taxonomic metadata for each image has been systematically integrated into the dataset. The lack of pre-annotated datasets has significantly restricted large-scale ecological analysis and the development of automated techniques in avian research, hindering the progress of AI-driven solutions tailored for bird species recognition. By addressing this gap, this dataset serves as a comprehensive benchmark for avian studies, fostering advancements in various applications such as conservation initiatives, environmental impact assessments, biodiversity preservation strategies, real-time tracking systems, and video-based analysis. Additionally, IBERBIRDS constitutes a resource for computer vision applications, supporting educational programs tailored to ornithologists and birdwatching communities. By openly providing this dataset, IBERBIRDS promotes scientific collaboration and technological advancements, ultimately contributing to the preservation and understanding of avian biodiversity.

PMID:40470349 | PMC:PMC12136710 | DOI:10.1016/j.dib.2025.111610

Data Brief. 2025 Apr 28;60:111593. doi: 10.1016/j.dib.2025.111593. eCollection 2025 Jun.

ABSTRACT

This dataset is an expanded version of a previously published collection of high-resolution RGB images of Urochloa spp. genotypes, initially designed to facilitate automated classification of phenological stages and raceme identification in forage breeding trials. The original dataset included 2400 images of 200 genotypes captured under controlled conditions, supporting the development of computer vision models for High-Throughput Phenotyping (HTP). In this updated release, 139 additional images and 24,983 new annotations have been added, bringing the dataset to a total of 2539 images and 47,323 raceme annotations. This version introduces increased diversity in image-capture conditions, with data collected from two geographic locations (Palmira, Colombia, and Ocozocoautla de Espinosa, Mexico) and a range of image-capture devices, including smartphones (e.g. Realme C53 and Oppo Reno 11), a Nikon D5600 camera, and a Phantom 4 Pro V2 drone. Images now vary in perspective (nadir, high-angle, and frontal) and capture distance (1-3 meters), enhancing the dataset applicability for robust Deep Learning (DL) models. Compared to the original dataset, raceme density per plant has nearly doubled in some samples, offering higher raceme overlap for advanced instance segmentation tasks. This expanded dataset supports deeper exploration of phenotypic variation in Urochloa spp. and offers greater potential for developing adaptable models in crop phenotyping.

PMID:40470345 | PMC:PMC12136700 | DOI:10.1016/j.dib.2025.111593

Data Brief. 2025 Apr 28;60:111587. doi: 10.1016/j.dib.2025.111587. eCollection 2025 Jun.

ABSTRACT

Breast cancer is one of the most prevalent cancers affecting women worldwide. Early detection and treatment are crucial in significantly reducing mortality rates Microcalcifications (MCs) are of particular importance among the various breast lesions. These tiny calcium deposits within breast tissue are present in approximately 30% of malignant tumors and can serve as critical indirect indicators of early-stage breast cancer. Three or more MCs within an area of 1 cm² are considered a Microcalcification Cluster (MCC) and assigned a BI-RADS category 4, indicating a suspicion of malignancy. Mammography is the most used technique for breast cancer detection. Approximately one in two mammograms showing MCCs is confirmed as cancerous through biopsy. MCCs are challenging to detect, even for experienced radiologists, underscoring the need for computer-aided detection tools such as Convolutional Neural Networks (CNNs). CNNs require large amounts of domain-specific data with consistent resolutions for effective training. However, most publicly available mammogram datasets either lack resolution information or are compiled from heterogeneous sources. Additionally, MCCs are often either unlabeled or sparsely represented in these datasets, limiting their utility for training CNNs. In this dataset, we present the MEXBreast, an annotated MCCs Mexican digital mammogram database, containing images from resolutions of 50, 70, and 100 microns. MEXBreast aims to support the training, validation, and testing of deep learning CNNs.

PMID:40470344 | PMC:PMC12136707 | DOI:10.1016/j.dib.2025.111587

Comput Struct Biotechnol J. 2025 Apr 29;27:1809-1817. doi: 10.1016/j.csbj.2025.04.034. eCollection 2025.

ABSTRACT

BACKGROUND AND OBJECTIVE: In recent years, DNA methylation-tumor classification based on artificial intelligence algorithms has led to a notable improvement in diagnostic accuracy compared to traditional machine learning methods. In cancer, the methylation pattern likely reflects both the cell of origin and somatically acquired DNA methylation changes, making this epigenetic modification an ideal tool for tumor classification. We propose an in-depth method based on the Convolutional Neural Network for the DNA methylation-based classification of papillary thyroid carcinoma (PTC) and its follicular (fvPTC) and classical (cvPTC) subtypes.

METHODS: To address this issue, we first performed a pan-cancer analysis to train a convolutional 1-D Neural Network (CNN) using supervised learning. Then, we evaluated the robustness of the net on an independent PTC dataset and assessed its ability to classify normal (N=56) versus tumor (N=461) samples and fvPTC (N=102) versus cvPTC (N=359). We then compared its performance with 4 machine learning models (logistic regression with elastic net penalty, quadratic discriminant analysis, support vector classifier with RBF kernel, and random forest).

RESULTS: By using RELU activation function and leaving out liquid tumors, our results show a remarkable performance of the neural network in classifying cancer and normal samples when applied to pan-cancer data (Validation AUC = 0.9903 and Validation Loss = 0.112). When applied to the thyroid independent dataset, the proposed Neural Net architecture successfully discriminates tumor versus normal samples (AUC = 0.91 +/- 0.05) and follicular versus classical PTC subtypes (AUC = 0.80 +/- 0.05), outperforming traditional machine learning algorithms.

CONCLUSIONS: In conclusion, the study highlights the effectiveness of CNNs in the methylation based classification of thyroid tumors and their subtypes, demonstrating its ability to capture subtle epigenetic differences with minimal preprocessing.This versatility makes the model adaptable for classifying other tumor types. Also, the findings emphasize the potential relevance of AI algorithms in addressing complex diagnostic challenges and supporting clinical decisions.This research lays the foundation for developing robust and generalizable models that can advance precision oncology in cancer diagnostics.

PMID:40470317 | PMC:PMC12136774 | DOI:10.1016/j.csbj.2025.04.034

Comput Struct Biotechnol J. 2025 May 1;27:1827-1837. doi: 10.1016/j.csbj.2025.04.032. eCollection 2025.

ABSTRACT

The increasing prevalence of multidrug-resistant bacteria, particularly Klebsiella species, poses a significant global health threat. Bacteriophages have emerged as promising alternatives due to their specificity and efficacy against bacterial targets. Characterizing phages, alongside analyzing their protein structures provide crucial insights into their host specificity, infection mechanisms, and potential applications. In this study, we isolated a novel bacteriophage, KPP105, and conducted comprehensive physiological, genomic, and structural analysis. Physiological assessments revealed that KPP105 maintains stable activity across a wide range of pHs and temperature conditions and exhibits host-specific infection properties. Genomic analysis classified KPP105 as a member of the Demerecviridae family and identified it as a lytic bacteriophage harboring a lytic cassette. Deep learning-based structural analysis of host-interacting proteins, including the receptor-binding protein (RBP) and endolysin derived from KPP105, was performed. Structural similarity analysis indicated that its RBP facilitates interactions with host receptors and exhibits unique sequence patterns distinguishing Klebsiella strains from other bacteria. Structure-based functional analysis provided comprehensive insights into cell wall degradation with various peptidoglycan fragments. In conclusion, this study reports the physiological, genomic, and structural characteristics of the novel lytic bacteriophage KPP105, offering valuable insights into its potential as an alternative agent against multidrug-resistant Klebsiella infections.

PMID:40470315 | PMC:PMC12136712 | DOI:10.1016/j.csbj.2025.04.032

Front Neurosci. 2025 May 21;19:1591410. doi: 10.3389/fnins.2025.1591410. eCollection 2025.

ABSTRACT

BACKGROUNDS: This study innovatively enhances personalized emotional responses and user experience quality in traditional Chinese folding armchair (Jiaoyi chair) design through an interdisciplinary methodology.

GOAL: To systematically extract user emotional characteristics, we developed a hybrid research framework integrating web-behavior data mining.

METHODS: 1) the KJ method combined with semantic crawlers extracts emotional descriptors from multi-source social data; 2) expert evaluation and fuzzy comprehensive assessment reduce feature dimensionality; 3) random forest and K-prototype clustering identify three core emotional preference factors: "Flexible Refinement," "Uncompromising Quality," and "ergonomic stability."

DISCUSSION: A CNN-GRU-Attention hybrid deep learning model was constructed, incorporating dynamic convolutional kernels and gated residual connections to address feature degradation in long-term semantic sequences. Experimental validation demonstrated the superior performance of our model in three chair design preference prediction tasks (RMSE = 0.038953, 0.066123, 0.0069777), outperforming benchmarks (CNN, SVM, LSTM). Based on the top-ranked preference encoding, we designed a new Jiaoyi chair prototype, achieving significantly reduced prediction errors in final user testing (RMSE = 0.0034127, 0.0026915, 0.0035955).

CONCLUSION: This research establishes a quantifiable intelligent design paradigm for modernizing cultural heritage through computational design.

PMID:40470295 | PMC:PMC12133947 | DOI:10.3389/fnins.2025.1591410

Res Sq [Preprint]. 2025 May 22:rs.3.rs-6641557. doi: 10.21203/rs.3.rs-6641557/v1.

ABSTRACT

MRI brain scans alone are not a definitive measure of dementia. Deep-learning algorithms (DLA) and professional human opinion are necessary for diagnosis. Yet, sample sizes are prohibitively large to train a typical DLA, which itself takes considerable computation time to produce diagnostically useful information from contrasting image features. We introduce analytic simplifications to this process to speed it up and reduce data requirements by modeling individual images as solutions of spatially autoregressive (AR) partial difference equations. Image features are the unique individual image AR parameters. Spatially lagged image pixels are explanatory variables for estimating a random-field representation (RFR) of the proposed AR difference equation. RFR model parameters are also those of the image autocorrelation function (ACF). An image pixel matrix-to-vector transformation allows AR parameters to be estimated by ordinary least squares (OLS) regression in millisecond time. Regression degrees of freedom (DOF) -- the number of image pixels -- are unusually large, leading to remarkably precise estimates of AR model parameters. These estimates support a solution of the binary dementia-normal classification of MRI axial brain scans (ADNI and OASIS archives). They also support the AR-RFR process applied to an original microscopic image of a glioblastoma cancer cell. In the face of formidable noise, a sharply defined and robust cancer cell model is estimated, which is an essential tool for cancer - type discrimination exercises and is parametrically plastic enough to serve a wide range of cells.

PMID:40470222 | PMC:PMC12136736 | DOI:10.21203/rs.3.rs-6641557/v1

Sci Rep. 2025 Jun 4;15(1):19532. doi: 10.1038/s41598-025-03765-3.

ABSTRACT

Current multi-stream convolutional neural network (MSCNN) exhibits notable limitations in path cooperation, feature fusion, and resource utilization when handling complex tasks. To enhance MSCNN's feature extraction ability, computational efficiency, and model robustness, this study conducts an in-depth investigation of these architectural deficiencies and proposes corresponding improvements. At present, there are some problems in multi-path architecture, such as isolated information among paths, low efficiency of feature fusion mechanism, and high computational complexity. These issues lead to insufficient performance of the model in robustness indicators such as noise resistance, occlusion sensitivity, and resistance to sample attacks. The architecture also faces challenges in data scalability efficiency and resource scalability requirements. Therefore, this study proposes an optimized model based on a dynamic path cooperation mechanism and lightweight design, innovatively introducing a path attention mechanism and feature-sharing module to enhance information interaction between paths. Self-attention fusion method is adopted to improve the efficiency of feature fusion. At the same time, by combining path selection and model pruning technology, the effective balance between model performance and computational resources demand is realized. The study employs three datasets, Canadian Institute for Advanced Research-10 (CIFAR-10), ImageNet, and Custom Dataset for performance comparison and simulation. The results show that the proposed optimized model is superior to the current mainstream model in many indicators. For example, on the Medical Images dataset, the optimized model's noise robustness, occlusion sensitivity, and sample attack resistance are 0.931, 0.950, and 0.709, respectively. On E-commerce Data, the optimized model's data scalability efficiency reaches 0.969, and the resource scalability requirement is only 0.735, showing excellent task adaptability and resource utilization efficiency. Therefore, the study provides a critical reference for the optimization and practical application of MSCNN, contributing to the application research of deep learning in complex tasks.

PMID:40467835 | DOI:10.1038/s41598-025-03765-3

Sci Rep. 2025 Jun 4;15(1):19612. doi: 10.1038/s41598-025-04631-y.

ABSTRACT

Prenatal healthcare development requires accurate automated techniques for fetal ultrasound image segmentation. This approach allows standardized evaluation of fetal development by minimizing time-exhaustive processes that perform poorly due to human intervention. This research develops a segmentation framework through U-Net++ with ResNet backbone features which incorporates attention components for enhancing extraction of features in low contrast, noisy ultrasound data. The model leverages the nested skip connections of U-Net++ and the residual learning of ResNet-34 to achieve state-of-the-art segmentation accuracy. Evaluations of the developed model against the vast fetal ultrasound image collection yielded superior results by reaching 97.52% Dice coefficient as well as 95.15% Intersection over Union (IoU), and 3.91 mm Hausdorff distance. The pipeline integrated Grad-CAM++ allows explanations of the model decisions for clinical utility and trust enhancement. The explainability component enables medical professionals to study how the model functions, which creates clear and proven segmentation outputs for better overall reliability. The framework fills in the gap between AI automation and clinical interpretability by showing important areas which affect predictions. The research shows that deep learning combined with Explainable AI (XAI) operates to generate medical imaging solutions that achieve high accuracy. The proposed system demonstrates readiness for clinical workflows due to its ability to deliver a sophisticated prenatal diagnostic instrument that enhances healthcare results.

PMID:40467763 | DOI:10.1038/s41598-025-04631-y

Genome Biol. 2025 Jun 4;26(1):156. doi: 10.1186/s13059-025-03619-1.

ABSTRACT

BACKGROUND: Whole-genome sequencing (WGS) data has facilitated genome-wide identification of rare noncoding variants. However, elucidating these variants' associations with complex diseases remains challenging. A previous study utilized a deep-learning-based framework and reported a significant brain-related association signal of autism spectrum disorder (ASD) detected from de novo noncoding variants in the Simons Simplex Collection (SSC) WGS cohort.

RESULTS: We revisit the reported significant brain-related ASD association signal attributed to deep-learning and show that local GC content can capture similar association signals. We further show that the association signal appears driven by variants from male proband-female sibling pairs that are upstream of assigned genes. We then develop Expression Neighborhood Sequence Association Study (ENSAS), which utilizes gene expression correlations and sequence information, to more systematically identify phenotype-associated variant sets. Applying ENSAS to the same set of de novo variants, we identify gene expression-based neighborhoods showing significant ASD association signal, enriched for synapse-related gene ontology terms. For these top neighborhoods, we also identify chromatin state annotations of variants that are predictive of the proband-sibling local GC content differences.

CONCLUSIONS: Overall, our work simplifies a previously reported ASD signal and provides new insights into associations of noncoding de novo mutations in ASD. We also present a new analytical framework for understanding disease impact of de novo mutations, applicable to other phenotypes.

PMID:40468385 | DOI:10.1186/s13059-025-03619-1

Head Face Med. 2025 Jun 4;21(1):44. doi: 10.1186/s13005-025-00521-w.

ABSTRACT

BACKGROUND: We aimed to comprehensively analyze the application of machine learning (ML) in dentistry and oral surgery using bibliometric methods to identify research trends, hotspots, and future directions.

METHODS: Publications related to ML in dentistry and oral surgery published between 2010 and 2024 were retrieved from the Science Citation Index Expanded by the Web of Science Core Collection (WoSCC). A total of 2234 unique publications were identified after screening. Bibliometric analysis was performed using the VOSviewer and CiteSpace software, focusing on parameters such as the number of publications, countries, institutions, journals, co-cited references, and keyword bursts.

RESULTS: The number of publications increased significantly from 2018 to 2024. China and the United States were the leading countries in terms of number of publications and citation counts. Prominent institutions include Seoul National University, Sichuan University, and Charite Universitätsmedizin Berlin. Journals such as BMC Oral Health and the Journal of Dentistry have a large number of publications. Analysis of the co-cited references revealed clusters related to disease diagnosis and risk prediction, treatment planning, clinical decision support systems, and dental education. Keyword bursts indicate the evolution of research focus from traditional machine learning algorithms to deep learning algorithms and the emerging importance of multimodal data and foundation models.

CONCLUSION: ML has made remarkable progress in dentistry and oral surgery. Although clinicians can benefit from the application of ML models in their practice, they should conduct comprehensive clinical validations to ensure the accuracy and reliability of these models. Moreover, challenges, such as data availability and security, algorithmic biases, and "black-box models", must be addressed. Future research should focus on integrating multimodal data and leveraging foundation models to improve the accuracy of diagnosis, treatment planning, and educational tools in dentistry and oral surgery.

PMID:40468381 | DOI:10.1186/s13005-025-00521-w

BMC Med Inform Decis Mak. 2025 Jun 4;25(1):207. doi: 10.1186/s12911-025-03027-2.

ABSTRACT

The detection and classification of blood cells are important in diagnosing and monitoring a variety of blood-related illnesses, such as anemia, leukemia, and infection, all of which may cause significant mortality. Accurate blood cell identification has a high clinical relevance in these patients because this would help to prevent false-negative diagnosis and to treat them in a timely and effective manner, thus reducing their clinical impacts.Our research aims to automate the process and eliminate manual efforts in blood cell counting. While our primary focus is on detection and classification, the output generated by our approach can be useful for disease prediction. This follows a two-step approach, where YOLO-based detection is first performed to locate blood cells, followed by classification using a hybrid CNN model to ensure accurate identification. We conducted a thorough and extensive comparison with other state-of-the-art models, including MobileNetV2, ShuffleNetV2, and DarkNet, for blood cell detection and classification. In terms of real-time performance, YOLOv10 outperforms other object detection models with better detection rates and classification accuracy. But MobileNetV2 and ShuffleNetV2 are more computationally efficient, which becomes more appropriate for resource-constrained environments. In contrast, DarkNet outperformed in terms of feature extraction performance, and the fine blood cell type classification. Additionally, an annotated blood cell data set was generated for this study. A diverse set of blood cell images with fine-grained annotations is contained in this dataset to make it useful for deep learning models training and evaluation. Because the present dataset will be an important resource for researchers and developers working on automatic blood cell detection and classification systems, we will make it publicly available under the open-access nature in order to accelerate the collaboration and progress in this field.

PMID:40468312 | DOI:10.1186/s12911-025-03027-2

BMC Med Inform Decis Mak. 2025 Jun 4;25(1):206. doi: 10.1186/s12911-025-03047-y.

ABSTRACT

BACKGROUND: Deep learning models have shown considerable potential to improve diagnostic accuracy across medical fields. Although YOLACT has demonstrated real-time detection and segmentation in non-medical datasets, its application in medical settings remains underexplored. This study evaluated the performance of a YOLACT-derived Real-time Polyp Delineation Model (RTPoDeMo) for real-time use on prospectively recorded colonoscopy videos.

METHODS: Twelve combinations of architectures, including Mask-RCNN, YOLACT, and YOLACT++, paired with backbones such as ResNet50, ResNet101, and DarkNet53, were tested on 2,188 colonoscopy images with three image resolution sizes. Dataset preparation involved pre-processing and segmentation annotation, with optimized image augmentation.

RESULTS: RTPoDeMo, using YOLACT-ResNet50, achieved 72.3 mAP and 32.8 FPS for real-time instance segmentation based on COCO annotations. The model performed with a per-image accuracy of 99.59% (95% CI: [99.45 - 99.71%]), sensitivity of 90.63% (95% CI: [78.95 - 93.64%]), specificity of 99.95% (95% CI: [99.93 - 99.97%]) and a F1-score of 0.94 (95% CI: [0.87-0.98]). In validation, out of 36 polyps detected by experts, RTPoDeMo missed only one polyp, compared to six missed by senior endoscopists. The model demonstrated good agreement with experts, reflected by a Cohen's Kappa coefficient of 0.72 (95% CI: [0.54-1.00], p < 0.0001).

CONCLUSIONS: Our model provides new perspectives in the adaptation of YOLACT to the real-time delineation of colorectal polyps. In the future, it could improve the characterization of polyps to be resected during colonoscopy.

PMID:40468304 | DOI:10.1186/s12911-025-03047-y

Med Phys. 2025 Jun 4. doi: 10.1002/mp.17910. Online ahead of print.

ABSTRACT

BACKGROUND: The reconstruction of a computed tomography (CT) image can be compromised by artifacts, which, in many cases, reduce the diagnostic value of the image. These artifacts often result from missing or corrupt regions in the projection data, for example, by truncation, metal, or limited angle acquisitions.

PURPOSE: In this work, we introduce a novel deep learning-based framework, latent space reconstruction (LSR), which enables correction of various types of artifacts arising from missing or corrupted data.

METHODS: First, we train a generative neural network on uncorrupted CT images. After training, we iteratively search for the point in the latent space of this network that best matches the compromised projection data we measured. Once an optimal point is found, forward-projection of the generated CT image can be used to inpaint the corrupted or incomplete regions of the measured raw data.

RESULTS: We used LSR to correct for truncation and metal artifacts. For the truncation artifact correction, images corrected by LSR show effective artifact suppression within the field of measurement (FOM), alongside a substantial high-quality extension of the FOM compared to other methods. For the metal artifact correction, images corrected by LSR demonstrate effective artifact reduction, providing a clearer view of the surrounding tissues and anatomical details.

CONCLUSIONS: The results indicate that LSR is effective in correcting metal and truncation artifacts. Furthermore, the versatility of LSR allows its application to various other types of artifacts resulting from missing or corrupt data.

PMID:40468155 | DOI:10.1002/mp.17910

Cell Regen. 2025 Jun 5;14(1):21. doi: 10.1186/s13619-025-00239-9.

ABSTRACT

Mechanical properties of cells have been proposed as potential biophysical markers for cell phenotypes and functions since they are vital for maintaining biological activities. However, current approaches used to measure single-cell mechanics suffer from low throughput, high technical complexity, and stringent equipment requirements, which cannot satisfy the demand for large-scale cell sample testing. In this study, we proposed to evaluate cell stiffness at the single-cell level using deep learning. The image-based deep learning models could non-invasively predict the stiffness ranges of mesenchymal stem cells (MSCs) and macrophages in situ with high throughput and high sensitivity. We further applied the models to evaluate MSC functions including senescence, stemness, and immunomodulatory capacity as well as macrophage diversity in phenotypes and functions. Our image-based deep learning models provide potential techniques and perspectives for cell-based mechanobiology research and clinical translation.

PMID:40468050 | DOI:10.1186/s13619-025-00239-9

Nat Comput Sci. 2025 Jun 4. doi: 10.1038/s43588-025-00810-z. Online ahead of print.

ABSTRACT

Quantum many-body methods provide a systematic route to computing electronic properties of molecules and materials, but high computational costs restrict their use in large-scale applications. Owing to the complexity in many-electron wavefunctions, machine learning models capable of capturing fundamental many-body physics remain limited. Here we present a deep learning framework targeting the many-body Green's function, which unifies predictions of electronic properties in ground and excited states, while offering physical insights into many-electron correlation effects. By learning the many-body perturbation theory or coupled-cluster self-energy from mean-field features, our graph neural network achieves competitive performance in predicting one- and two-particle excitations and quantities derivable from a one-particle density matrix. We demonstrate its high data efficiency and good transferability across chemical species, system sizes, molecular conformations and correlation strengths in bond breaking, through multiple molecular and nanomaterial benchmarks. This work opens up opportunities for utilizing machine learning to solve many-electron problems.

PMID:40468046 | DOI:10.1038/s43588-025-00810-z

Eur Spine J. 2025 Jun 4. doi: 10.1007/s00586-025-08995-7. Online ahead of print.

NO ABSTRACT

PMID:40468028 | DOI:10.1007/s00586-025-08995-7