Synthetic aperture radar can produce two-dimensional observations regardless of daylight or weather, making it useful for maritime monitoring, but ship targets are difficult to distinguish from the surrounding sea and coast. Their radar signatures can appear as scattered bright points rather than continuous objects, while speckle noise blurs boundaries and distorts local texture. Detection is further complicated by an exceptionally wide target-size range and by settings ranging from isolated fishing boats in open water to crowded berths, islands, buildings and other coastal clutter. The study addressed these problems by modifying three parts of YOLOv8: how features are extracted, how useful features are selected during multi-scale fusion and how predicted boxes are matched to ship shapes.

Coordinate attention, or CA, was added throughout the network’s neck, where features from different spatial scales are combined. Unlike channel-only attention, which summarizes a feature map without retaining its position, CA pools information separately along the horizontal and vertical axes. It then applies learned weights that preserve coordinate cues while identifying relationships among feature channels. In the model, this mechanism was intended to strengthen valid ship responses and suppress sea-surface speckle, nearshore structures and other distracting patterns. Ablation tests showed why placement mattered. Adding CA to the backbone produced precision of 0.878 and recall of 0.874, both below the baseline, while placing it in the neck yielded 0.891 precision and 0.923 mAP@0.5. The neck-embedded version also improved mAP@0.5:0.95 from 0.525 to 0.531 without changing the reported 909-frame-per-second throughput.

The second modification replaced one standard convolution in the C2f feature blocks at the P4 and P5 backbone stages with dynamic snake convolution, or DSConv. Standard convolution samples a fixed rectangular grid, which can poorly follow the curved or irregular outlines of vessels in radar imagery. DSConv first learns offsets that describe changing contour trends, then moves its sampling positions toward informative target regions. The researchers argued that this flexible pattern can preserve fine scattering details more effectively than fixed sampling while avoiding some background responses. Comparisons with four alternative convolution modules supported that interpretation. DSConv produced 0.901 precision, 0.899 recall and 0.929 mAP@0.5, outperforming the baseline and the SPDConv, WTConv and SAConv alternatives on conventional detection accuracy. Its strict localization score of 0.495 was nevertheless below the baseline’s 0.525, and its reported throughput was 476 frames per second.

The third component, Shape-IoU, replaced the baseline’s complete intersection-over-union branch for bounding-box regression. Shape-IoU combines overlap, scale deviation, center distance and explicit aspect-ratio consistency, with weights derived from the width and height of each ground-truth box. The intended effect is to favor predicted boxes that resemble the geometry of actual ships rather than rewarding overlap alone. Used independently, the loss raised precision to 0.895, recall to 0.897, mAP@0.5 to 0.919 and mAP@0.5:0.95 to 0.531. The authors noted, however, that ships with blurred or low-contrast boundaries can have ambiguous physical outlines. Strong shape constraints may improve detection at a conventional IoU threshold while causing small boundary offsets that are penalized by the stricter 0.5-to-0.95 metric.

Joint ablation experiments exposed an interaction among the three optimizations that was not evident from their individual results. Coordinate attention alone, DSConv alone and Shape-IoU alone each maintained or improved strict localization performance. Combining CA and DSConv produced the study’s highest mAP@0.5:0.95, 0.532, but adding Shape-IoU reduced that value to 0.514. The complete configuration nevertheless produced the best precision, recall and mAP@0.5, reaching 0.908, 0.904 and 0.925. The authors describe this as a multi-module coupling effect and suggest that dynamic sampling, feature re-weighting and shape regularization may accumulate small boundary perturbations for speckle-blurred targets. Those perturbations have limited consequences at IoU 0.5 but carry greater weight when evaluation demands close pixel-level alignment across a range of increasingly strict thresholds.

The model was trained and tested on 39,729 chips generated from 210 SAR scenes in the updated 2021 SAR-SHIP dataset. The source scenes comprised 72 images from China’s Gaofen-3 satellite and 138 from the European Space Agency’s Sentinel-1 satellite. Spatial resolutions ranged from 1.7 to 25 metres, and the collection covered HH, HV, VH and VV polarizations in port, nearshore, island and open-sea environments. After cropping and deduplication, the chips were randomly divided at the image-chip level into training, validation and test sets containing 27,902, 5,980 and 5,980 images, respectively. Researchers removed geographic overlap and duplicate crops to reduce leakage, manually corrected ambiguous annotations and resized all inputs to 320 by 320 pixels. This dataset design increases geographic diversity, although chip-level random partitioning cannot represent every possible form of overlap between scenes.

Training used an NVIDIA RTX 4090 with 24 gigabytes of memory, PyTorch 2.5.0 and CUDA 12.1. All models ran for 200 epochs with a batch size of 16 and an initial learning rate of 0.0015. Precision and recall were calculated at a confidence threshold of 0.25, while mAP values were aggregated across confidence thresholds under the COCO evaluation protocol. The researchers compared the proposed detector with YOLOv5, YOLOv8, YOLOv10, YOLOv11, YOLOv12 and RT-DETR-L. At conventional thresholds, the improved model delivered the strongest reported results: its 0.925 mAP@0.5 exceeded YOLOv12’s 0.923, YOLOv10’s and YOLOv5’s 0.921, and YOLOv11’s 0.918. It also outperformed RT-DETR-L in precision, recall and mAP@0.5, while using 11.19 million parameters rather than the transformer model’s 31.99 million.

The computational cost remained close to that of the baseline: the improved model contained 11.19 million parameters and required 28.6 billion floating-point operations, compared with 11.126 million and 28.4 billion for YOLOv8. Runtime, however, was less favorable because offset prediction, dynamic sampling and feature interpolation add memory-access-heavy operations not fully represented by the theoretical FLOP count. Such irregular operations also benefit less from the highly optimized routines used for regular GPU convolution. The reported 588 frames per second therefore reflected only forward-pass inference, excluding image loading, preprocessing and postprocessing, and the authors treated throughput as a secondary indicator sensitive to hardware and system load. The results support the model for accuracy-oriented applications with moderate real-time requirements, not for systems requiring the baseline’s maximum processing speed or the precise localization captured by mAP@0.5:0.95.

Cross-dataset tests indicated that the improvements were not confined to the training source. Weights trained on SAR-SHIP 2021 were applied directly, without reported retraining, to the HRSID and SSDD datasets at the same 320-pixel input size. On HRSID, the improved model raised precision from 0.824 to 0.832, recall from 0.507 to 0.508 and mAP@0.5 from 0.544 to 0.547, while mAP@0.5:0.95 fell from 0.213 to 0.204. On SSDD, precision increased from 0.816 to 0.825, recall from 0.634 to 0.646 and mAP@0.5 from 0.684 to 0.692, with strict localization declining from 0.263 to 0.247. Visual examples across four polarization modes also showed more confident detections and fewer clutter-induced false alarms than the baseline. These findings support favorable transfer across differing SAR data distributions, but the experiment did not establish performance for every radar sensor, sea state or operating environment.

Subject of Research: Multi-module YOLOv8 optimization for ship detection in high-resolution synthetic aperture radar imagery

Article Title: Multi-scale ship detection with high-resolution SAR images based on improved YOLOv8

Article References: Liu, D., Chen, H., Chen, J., Ding, Y., Zhu, J., & Lei, P. (2026). Multi-scale ship detection with high-resolution SAR images based on improved YOLOv8. Frontiers in Marine Science, 13, Article 1943050. https://doi.org/10.3389/fmars.2026.1943050

Image Credits: AI Generated

DOI: 10.3389/fmars.2026.1943050

Keywords: synthetic aperture radar, ship detection, YOLOv8, coordinate attention, dynamic snake convolution, Shape-IoU, remote sensing, computer vision, object detection, cross-dataset generalization, Multi-scale, ship

The investigation focused on perirenal adipose tissue, a visceral fat depot enclosed by Gerota’s fascia and characterized by rich vascularization and immune-cell infiltration. Tissue next to a solid tumor can undergo metabolic and inflammatory remodeling, including changes in adipocytes, extracellular matrix and vascular tissue that may alter its CT attenuation. Adipose tissue is ordinarily associated with negative Hounsfield-unit values, so a shift toward zero can reflect changes in lipid content, fibrosis or inflammation. The researchers examined whether this imaging signal varied with tumor position, size and pathological features, and whether the same measurements could be reproduced with artificial intelligence rather than relying exclusively on hand-selected regions of interest. The goal was quantitative characterization of an imaging marker, not a test of whether perirenal fat changes cause cancer or affect outcomes.

All cancer participants had non-enhanced computed tomography scans obtained between May 2015 and December 2021 at a single Turkish institution. Of 118 consecutive patients initially screened, 61 were excluded because of motion artifacts, technical limitations or inadequate image quality. The final cancer cohort had a mean age of 65.1 years, compared with 50.0 years among controls, a statistically significant difference. Controls were evenly divided between 41 men and 41 women and distributed across age decades from 30 to 69, but the groups as a whole were not age-matched. A senior radiologist reviewed the images and placed ten circular regions of interest per kidney, each measuring 50 to 70 square millimeters, at the upper, middle and lower poles while avoiding vessels, fascia, tumor margins and artifacts. Images were acquired with 1-mm slices on a 128-row multidetector scanner, using tube voltages of 100 to 120 kV and body-mass-index-modulated tube currents.

At the upper poles, mean perirenal fat density was −97.64 ± 10.73 HU on the right in the cancer group, compared with −94.99 ± 9.76 HU in controls, and −100.39 ± 10.20 HU on the left, compared with −96.55 ± 10.57 HU in controls. The corresponding p values were 0.040 and 0.032. No significant case-control differences were detected at either middle or lower pole. Within controls, right and left measurements differed significantly at all three levels, showing that perirenal fat density has a degree of anatomical asymmetry even in people without known renal or systemic disease. In the cancer group, a significant right-left difference was confined to the upper pole, where the p value was 0.005. This side and region dependence means that a single average measurement could obscure differences present in individual anatomical areas.

Tumor geometry helped refine the spatial pattern. When each of the six measurements was classified by its relationship to the tumor, fat directly adjacent to the tumor was denser than fat at non-adjacent poles on the same side, at the matching pole on the opposite side and at non-matching contralateral poles. Mean density was −93.2 ± 11.5 HU adjacent to the tumor, compared with −96.0 ± 16.6 HU in ipsilateral non-adjacent regions, −100.1 ± 7.7 HU in contralateral matched poles and −99.7 ± 8.4 HU in contralateral non-matched poles; paired comparisons gave p = 0.0025, p < 0.0001 and p < 0.0001, respectively. Tumor dimensions were also positively correlated with density in the ipsilateral upper pole, including correlations of 0.284 and 0.369 with total anteroposterior and transverse dimensions. These associations support a localized tumor–fat interaction, but they do not establish that fat-density changes cause tumor growth or invasion.

The upper-pole case-control pattern was not confined to tumors located at the top of the kidney. Among 17 patients with lower-pole tumors, upper-pole fat density was lower than in controls on both sides, with p values of 0.032 and 0.018. Patients with middle-zone tumors did not show the same differences, and comparisons among tumor-location groups within the cancer cohort were not significant. Other findings were similarly selective: density did not differ significantly by sex or between the 41 clear-cell and 12 papillary renal cell carcinomas. The 12 papillary tumors included three WHO/ISUP type 1 and nine type 2 cancers, which were analyzed together because the subtype-specific groups were too small for a reliable comparison. Density also was not significantly associated with lymphovascular invasion, Gerota’s fascia thickening or tumor necrosis, although it was consistently less negative in the 13 lymphovascular-invasion-positive cases.

Larger tumor volume, however, was strongly associated with necrosis. No necrosis was observed among 13 tumors smaller than 17 cc, compared with 11 of 20 tumors in the 40.01–134 cc group and 7 of 13 tumors above 134 cc; Fisher’s exact test gave p = 0.003. Discrimination between cancer and control groups was modest and depended on measurement location. The area under the receiver operating characteristic curve was 0.603, with a 95% confidence interval of 0.505–0.700, for the right upper pole and 0.607, with an interval of 0.511–0.704, for the left upper pole. These were the only regions whose confidence intervals excluded an AUC of 0.5. Values at the right and left middle poles were 0.561 and 0.464, while lower-pole values were 0.416 and 0.444. The design also compared confirmed cancers with healthy controls rather than patients with indeterminate or benign renal masses, so it did not test the discrimination needed to establish clinical diagnostic value.

For automated analysis, the team trained a YOLOv8x segmentation model on 1,637 CT slices from the 139 participants, with about 12 slices per patient spanning the upper, middle and lower kidney regions. A senior radiologist and anatomist manually annotated the left and right perirenal-fat regions. Images were resized to 640 × 640 pixels, and the model was trained for 100 epochs with stochastic gradient descent and a learning rate of 0.01 on an Nvidia A100 graphics processor. Data were divided by patient, rather than by slice, to prevent information leakage between training and evaluation sets, with 70% used for training, 20% for validation and 10% for testing. At the best epoch, precision was 79.6%, recall 78.8% and mean average precision at an intersection-over-union threshold of 0.5 was 82.98%. These metrics describe performance within the study dataset; no external test set or cross-validation was used.

The automated density estimates agreed best with the radiologist’s measurements when the analysis included CT values from −125 to +30 HU. Wider normalization windows produced much lower density estimates: for example, the mean right-upper-pole value in cases was −132.71 ± 3.39 HU with the −150 to −50 HU interval and −162.44 ± 5.41 HU with the −190 to −30 HU interval, compared with −97.64 ± 10.73 HU manually. Even the closest interval produced statistically significant differences in several subgroup-level comparisons, including the left upper pole in both cases and controls. The result shows that an AI segmentation system does not eliminate measurement-definition problems: threshold selection materially changes the derived biomarker. The model also required manually selected and annotated regions during development and validation, and its ability to transfer to different hospitals, scanners and patient populations remains untested.

The largest threat to the headline case-control finding is age. An exploratory multivariable regression accounting for group, age and sex found that the upper-pole difference was no longer significant after age adjustment, with p = 0.51 on the right and p = 0.56 on the left. Age was an independent inverse predictor of density at p < 0.001 for both sides of upper-pole PRFD, but the reported age relationship differed between groups: within controls, age correlated with right upper-pole density at r = −0.535, p < 0.001, whereas the corresponding correlation was not significant within cases at r = −0.038, p = 0.78. The retrospective design, modest sample, single-center setting and small subgroups—including 12 papillary cancers and 13 cancers with lymphovascular invasion—further restrict interpretation. The results identify a spatially heterogeneous imaging feature and a promising computational workflow, but not a clinically validated biomarker for detecting or characterizing renal cell carcinoma.

Subject of Research: Regional CT measurement of perirenal fat density in renal cell carcinoma and automated segmentation with YOLOv8

Article Title: Quantitative assessment of perirenal fat density in renal cell carcinoma: a combined radiologist-driven and YOLO-based CT study

Article References: Sasani, H., Özkan, M., Şimşek, M. A., Akgül, H. M., & İsal Arslan, A. (2026). Quantitative assessment of perirenal fat density in renal cell carcinoma: a combined radiologist-driven and YOLO-based CT study. Frontiers in Radiology, 6, 1911034. https://doi.org/10.3389/fradi.2026.1911034

Image Credits: AI Generated

DOI: 10.3389/fradi.2026.1911034

Keywords: renal cell carcinoma, perirenal fat, computed tomography, Hounsfield units, YOLOv8, medical imaging, image segmentation, age confounding, tumor microenvironment, Quantitative, assessment, perirenal

Its primary target was the posterior primary dentition, specifically the first primary molars, designated D, and second primary molars, designated E. The authors describe the observed performance as encouraging but caution that larger, more diverse and multicentre datasets will be needed to determine how robust and generalizable the pipeline is, particularly in the presence of potentially confounding restorative materials. The first stage used YOLOv8 to locate the target teeth in each photograph.

The 725 images were divided into training, validation and test sets at a ratio of 70%, 15% and 15%, yielding 508, 109 and 108 photographs, respectively. The held-out test set contained 303 annotated teeth. On this set, the detector achieved precision of 0.963, recall of 0.972 and mean average precision at an intersection-over-union threshold of 0.5, or mAP@0.5, of 0.983.

Detected teeth then entered a segmentation stage based on the Segment Anything Model, abbreviated SAM, which isolated the tooth region for classification. A ResNet50 network was used to classify each segmented tooth image into one of the five predefined categories. The analysis separated the pipeline into these linked stages so that the system could first identify individual posterior teeth and then assess the visible condition represented in each tooth’s image.

For the main primary tooth-level classification analysis, the researchers divided 1,946 tooth samples into 1,362 training, 389 validation and 195 test samples, corresponding to a 70%, 20% and 10% split. The ResNet50 classifier achieved 87.7% accuracy on the 195 held-out teeth. In this analysis, teeth originating from the same source photograph were not necessarily restricted to one subset of the data, a design detail that the investigators addressed with an additional sensitivity analysis.

That sensitivity analysis used a 70%, 15% and 15% division of the 725 source photographs, maintaining separation at the patient or source-photograph level. An independently retrained ResNet50 was then evaluated on 290 held-out tooth samples. It achieved 85.9% accuracy, 85.6% macro-averaged precision, 85.7% macro-averaged recall and an 85.6% F1-score.

Agreement between classifications was also substantial, with a Cohen’s κ of 0.822. The lower accuracy than in the original tooth-level test did not erase the overall performance, but it quantified how results changed under the alternative grouping strategy. The researchers also applied Grad-CAM as a qualitative visualization tool.

This technique produced heat maps identifying image regions that contributed to the classifier’s predictions, allowing the team to examine which parts of the segmented tooth images influenced the assigned categories. The visualization was interpretive rather than a separate quantitative test: it showed the features associated with model decisions, but it did not by itself establish diagnostic accuracy or clinical usefulness. Overall, the study provides image-based evidence that a multistage deep-learning pipeline can combine tooth localization, segmentation and classification to assess primary molars in pediatric intraoral photographs.

Its strongest reported results were 0.963 precision and 0.972 recall for detecting D and E teeth, together with 87.7% classification accuracy in the original held-out tooth analysis and 85.9% in the source-photograph-level sensitivity analysis. The authors frame the approach as a potential aid for standardized photographic assessment and workflow efficiency in pediatric dentistry, while emphasizing that broader external validation is still required before its robustness can be assessed beyond the study’s dataset.

Subject of Research: Evaluation of a multistage deep-learning pipeline for detecting and classifying caries and restorations in pediatric primary molars from intraoral photographs.

Article Title: Artificial Intelligence-based multi-stage pipeline for detecting caries and restorations in pediatric primary molars from intraoral photographs

Article References: Kavousinejad, S., Motie, P., Daghighi, Y., Norouzi, A., Rabbani, H., Biglar, N., Ebadifar, A., & Behnaz, M. (2026). Artificial Intelligence-based multi-stage pipeline for detecting caries and restorations in pediatric primary molars from intraoral photographs. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02975-z

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02975-z

Keywords: artificial intelligence, pediatric dentistry, primary molars, dental caries, intraoral photographs, deep learning, YOLOv8, image segmentation, ResNet50, computer-assisted diagnosis, Artificial, Intelligence-based

This was a computational model-development and benchmarking study rather than a review or clinical investigation. Its scope included pig-face images from the PetFace pig subset and the FIRC dataset, which the authors divided into training, validation and test sets at a ratio of 7:2:1. Data augmentation included random flipping, random cropping and color-gamut transformation to represent variations such as backlighting, low light, occlusion and changes in posture. All comparison models used fivefold cross-validation, and the researchers used an independent-samples t test to assess differences between groups.

On the PetFace data, I-YOLOv8-MCGIRNet recorded precision of 0.956, recall of 0.948 and mAP@0.5:0.95 of 0.897. The reported p-values for comparisons with the other models were below 0.05. In separate performance-and-efficiency comparisons, the model used 15 million parameters and produced mAP@0.5 of 0.966 on PetFace. Comparative results reported for that dataset included approximately 0.933 for YOLOv8s-MCGIRNet, 0.947 for I-YOLOv8-Swin Transformer and 0.885 for Faster R-CNN, whose model size was about 68 million parameters.

The FIRC results showed a similar pattern. I-YOLOv8-MCGIRNet achieved mAP@0.5 of 0.955 with 15 million parameters. The reported mAP@0.5 values were 0.924 for YOLOv8s-MCGIRNet, 0.937 for I-YOLOv8-Swin Transformer and 0.875 for Faster R-CNN. An ablation experiment attributed successive gains to the global attention module, BiFPN feature fusion, CIoU loss and MCGIRNet. Embedding MCGIRNet increased mAP@0.5 to 0.966 while reducing the model to 15 million parameters, according to the study.

Under the researchers’ definition of recognition consistency, the model’s value increased from about 90.9% to 94.5% in normal lighting, from 87.9% to 94.8% in dark light below 200 lux and from 85.4% to 92.7% under backlighting with a light-intensity difference above 800 lux. The computational tests reported 16.32 GFLOPs for I-YOLOv8-MCGIRNet, compared with 19.68 for YOLOv8s-MCGIRNet, 22.75 for I-YOLOv8-IRNet, 47.21 for I-YOLOv8-Swin Transformer and 118.46 for Faster R-CNN. Peak GPU memory was 8.26 GB on PetFace and 7.94 GB on FIRC, while reported epoch times were 4.12 and 3.28 minutes, respectively.

Evaluation in large-scale pig-farm settings covered nursery, fattening and breeding houses. Each scene type included 10 breeding stalls, with pig images collected over 30 consecutive days under varying lighting, postures and stocking densities. I-YOLOv8-MCGIRNet was reported to detect one pig face in 2.8 milliseconds. Its small-target detection rate in dense scenes was 94.1%, compared with 86.3% for I-YOLOv8-IRNet. When 30% of local facial features were partially missing, its recognition accuracy was 82.7%, versus 79.6% for I-YOLOv8-Swin Transformer and 61.8% for Faster R-CNN.

Counting performance was measured in sparse, dense and overlapping scenarios. For I-YOLOv8-MCGIRNet, counting accuracy rose from 88.6% to 98.5% in sparse scenes, from 79.9% to 94.7% in dense scenes and from approximately 71.8% to 92.2% in overlapping scenes. In the overlapping condition, Faster R-CNN rose from about 52.1% to 64.9%. These results apply to the sampled farm environments and test conditions used in the study rather than establishing performance across every type of pig facility.

Recognition was also assessed with pigs standing, lying down and turning their heads 45 degrees left or right. The study reported that I-YOLOv8-MCGIRNet achieved the best detection and recognition accuracies among the compared models for all three postures, with both measures remaining above 0.91 in the non-standard postures. In the head-turning condition, its detection and recognition accuracies were 0.938 and 0.919. Relative to I-YOLOv8-IRNet, the reported improvements were 7.0 and 9.4 percentage points, respectively.

The evidence supports the model’s measured performance within the selected datasets and farm scenarios, but its generalizability remains untested. The authors trained, validated and tested within PetFace and FIRC separately and did not perform direct cross-dataset transfer validation. The study also did not fully evaluate outdoor conditions in extreme weather or optimize the system specifically for facial differences among pig breeds and ages. The results therefore indicate benchmark and scenario-specific performance for pig-face recognition and counting, rather than demonstrated effectiveness in unstudied farms, species, breeds or environmental conditions.

Subject of Research: Experimental development and benchmarking of a pig-face recognition and counting model

Article Title: Intelligent recognition and counting of pig faces based on improved YOLOv8 and MCGIRNet

Article References: Kang, J., Wang, H., Chen, C., & Liu, Q. (2026). Intelligent recognition and counting of pig faces based on improved YOLOv8 and MCGIRNet. Discover Artificial Intelligence, 6(1), Article 1467. https://doi.org/10.1007/s44163-026-02349-y

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02349-y

Keywords: pig-face recognition, pig counting, YOLOv8, MCGIRNet, computer vision, object detection, precision livestock, smart farming, machine learning, Intelligent, recognition, counting

Published in the journal SoftwareX, the new tool represents a significant step forward in digital plant pathology. Developed by Stefanie Lück, Moozhan Serpoush, and Dimitar Douchkov, BluVision Root utilizes deep learning-based instance segmentation models to detect and quantify disease symptoms from standardized seedling images. Unlike previous image-analysis tools such as SmartRoot or RootNav, which primarily focus on root architecture traits like length and branching, BluVision Root is specifically engineered to identify Fusarium-infected regions alongside total root tissue. This dual detection capability allows for the extraction of disease-specific quantitative traits, providing a continuous and objective measure of infection severity that can be directly applied to statistical and genetic analyses.

The technical architecture of BluVision Root is built upon the Ultralytics YOLO framework, a popular deep learning library for object detection and segmentation. The software processes composite RGB images, which are often generated from monochrome images acquired at defined wavelengths by the Macrobot high-throughput phenotyping platform. A central technical challenge in developing this tool was the reliable segmentation of thin root structures and relatively small infected tissue regions against a large image background. Because infected regions often occupy only a small fraction of the total image area, the model faced a significant small-object segmentation problem with substantial foreground-background imbalance. To overcome this, the developers designed the architecture to support high-resolution image processing and polygon-based annotation, ensuring that the model could accurately distinguish between healthy and diseased tissue even when the symptoms were subtle or localized.

Model development involved a rigorous process of manual annotation and systematic optimization. Researchers manually annotated 128 root images using the Roboflow platform, defining two biologically relevant object classes: total root tissue and Fusarium-infected regions. The training dataset contained 597 root annotations and 392 Fusarium-infected region annotations. During the development phase, the team evaluated 39 different model configurations, including variants of YOLOv8, YOLO11, and YOLO26, with various model sizes, image resolutions, and training durations. The models were compared using standard metrics such as mask mAP50-95, precision, and recall. Ultimately, the YOLOv8s-seg model, trained at an image resolution of 1024 pixels for 150 epochs using the AdamW optimizer, provided the best overall performance and was selected as the production model for the software release.

The validation of BluVision Root was conducted using an independent dataset of 148 samples from winter wheat seedlings inoculated with Fusarium culmorum strain UK99. These seedlings were imaged daily from one to four days post-infection using the Macrobot 2.0 platform under controlled environmental conditions. The automated infection-area measurements generated by the software were compared against expert-based manual scoring, where an expert visually estimated the percentage of total root tissue showing Fusarium-associated symptoms on a continuous scale from 0 to 100%. The results demonstrated strong agreement between the automated and manual assessments, with a Pearson correlation coefficient of 0.906. The mean absolute error was 9.51 percentage points, and the root mean square error was 13.35 percentage points, indicating that the software provides a reliable and accurate alternative to human scoring for quantitative disease assessment.

From a user perspective, BluVision Root offers a modular and flexible workflow that supports batch processing of experimental image sets, object-level detection, and quantitative trait extraction. The software calculates several disease-related traits, including infected root tissue area, total root tissue area, relative infection coverage, and symptom distribution. It also supports lane-position analysis, which is particularly useful for plate-based high-throughput experiments where seedlings are arranged in fixed positions. The output includes segmentation masks, polygon coordinates, class labels, and confidence values, allowing users to inspect the detected structures and export standardized CSV tables for downstream analysis. These tables can be used directly for statistical modeling, genotype comparison, treatment evaluation, or machine learning workflows, facilitating a seamless transition from image data to biological insight.

Performance benchmarks indicate that BluVision Root is capable of handling large datasets efficiently, even without GPU acceleration. In a CPU-only benchmark using 2,268 images on a Windows workstation equipped with two Intel Xeon E5-2643 v3 processors and 128 GB of RAM, the complete processing workflow required approximately 10,168 seconds. This corresponds to an average processing time of 4.48 seconds per image, or roughly 13.4 images per minute. While GPU acceleration can substantially speed up model inference, the CPU-only performance demonstrates that the software is accessible to laboratories with standard computing infrastructure. The modular architecture of the software also allows individual components, such as preprocessing steps, model weights, and reporting settings, to be updated or adapted without altering the complete software structure, ensuring long-term maintainability and adaptability to new research needs.

The impact of BluVision Root extends beyond simple disease quantification to enable research questions that are difficult to address with manual scoring alone. By providing continuous, reproducible phenotypic data, the software facilitates the analysis of early infection dynamics, genotype-specific differences in symptom development, and spatial patterns of infection along seedling tissues. This is particularly relevant for resistance research and breeding programs, where large genotype panels require rapid and objective assessment of disease responses. The extracted traits provide a quantitative basis for statistical modeling and biological interpretation, supporting applications such as QTL mapping, genome-wide association studies, and machine learning-based prediction of disease resistance. By reducing operator-dependent variation, the software improves the comparability of disease measurements across different experiments, time points, and research groups.

Despite its advantages, the current implementation of BluVision Root has defined limitations that researchers must consider. The model was validated specifically for young wheat seedlings inoculated with F. culmorum UK99 under standardized Macrobot imaging conditions. It does not establish a direct correspondence between seedling responses and resistance in soil-grown or adult plants. Additionally, non-pathogenic discoloration that resembles Fusarium-associated symptoms may lead to misclassification, making the use of non-inoculated controls advisable for accurate biological interpretation. Application to other crops, diseases, or substantially different imaging conditions may require representative training data, model retraining, and independent validation. The current release also does not automatically identify or exclude blurred, overexposed, or otherwise unsuitable images, necessitating a quality control step in the workflow.

BluVision Root is available as open-source software under the GNU General Public License v3.0, with a public repository and archived software release to support reuse and further development. By linking controlled infection assays with automated quantitative image analysis, the tool provides a practical and reusable resource for digital plant pathology, resistance genetics, and cereal breeding research. As the field of plant phenomics continues to evolve, tools like BluVision Root are essential for overcoming the phenotyping bottleneck and accelerating the discovery of disease-resistant crop varieties. The software’s ability to transform complex image data into actionable quantitative traits marks a significant advancement in the precision and efficiency of plant disease research.

Subject of Research: Automated image analysis software for high-throughput phenotyping of Fusarium root rot in cereal seedlings

Article Title: BluVision Root: automated image analysis software for high-throughput phenotyping of Fusarium root rot in cereal seedlings

Article References: Lück, S., Serpoush, M., & Douchkov, D. (2026). BluVision Root: automated image analysis software for high-throughput phenotyping of Fusarium root rot in cereal seedlings. SoftwareX, 36, Article 103132. https://doi.org/10.1016/j.softx.2026.103132

Image Credits: AI Generated

DOI: 10.1016/j.softx.2026.103132

Keywords: Fusarium root rot, Cereal phenotyping, Deep learning, YOLOv8, Plant pathology, Image analysis, Open-source software, High-throughput screening, Wheat disease, Digital agriculture, BluVision, Root

The study, published in npj Heritage Science, takes five traditional villages in Guangdong Province as its case studies. These settlements share the characteristics that make manual inspection so difficult: complex spatial layouts, dense clusters of traditional buildings, narrow lanes that resist ground-based survey, and locations remote enough that they rarely attract the public attention or funding streams that urban monuments receive. When inspection is slow, repairs are delayed, and delayed repairs accelerate decay in a feedback loop that has already erased so much of China’s rural built heritage.

The team’s answer begins in the air. Using unmanned aerial vehicles equipped with visible-light cameras, they acquired a total of 12,762 images over the five villages. These were not casual snapshots but systematically collected photographs designed for photogrammetry, the technique of reconstructing three-dimensional geometry from overlapping two-dimensional images. Processed together, the images yield high-precision spatial digital archives: detailed orthophoto maps and three-dimensional models that record the state of every visible roof, wall, and courtyard at the moment of capture. Such archives serve a dual purpose. They are preservation documents in their own right, preserving a measurable record of structures that may not survive, and they are the raw material for automated condition assessment.

On top of that digital foundation sits the deep learning component. The researchers trained an object detection model based on the YOLOv8 architecture, a family of convolutional neural networks designed to identify and localize objects in images in a single pass. To teach the model what roof damage looks like in traditional Guangdong villages, the team assembled a training dataset of 41,213 annotated instances, each one a labeled example of damage or a feature of interest drawn from the drone imagery. Annotation at that scale is laborious work, but it is what allows a neural network to distinguish, say, a missing tile or a collapsed roof section from the ordinary textures of aged clay and weathered timber.

The performance figures reported in the paper are notable. The YOLOv8-based model achieved an overall mean average precision at an intersection-over-union threshold of 0.5, abbreviated mAP50, of 0.832. In practical terms, mean average precision combines how accurately the model detects damaged regions with how precisely it draws the bounding boxes around them; a value above 0.8 indicates that the system correctly identifies the large majority of damage instances while keeping false alarms at a manageable level. For heritage managers deciding where to send repair crews, that balance matters enormously, because every false positive wastes a site visit and every false negative is a building left to decay.

Speed is the other half of the story. The researchers validated their trained model on ultra-high-resolution digital orthophoto maps, known as DOMs, and on aerial video footage captured over the villages. Across these validation runs, the model averaged an inference time of 5.77 milliseconds per frame. That figure means the system can process aerial video essentially in real time, scanning continuous footage from a drone flight and flagging damage as it appears, rather than requiring hours of offline computation. Combined with the orthophoto analysis, it allows an entire village to be surveyed from the air in a fraction of the time a ground-based inspection would demand.

The technical logic of the workflow is straightforward to describe but demanding to execute. Photogrammetry converts the 12,762 overlapping drone images into georeferenced orthophotos and three-dimensional models, ensuring that every detected damage instance can be tied to a real-world location on a specific building. The deep learning model then sweeps across these products, drawing bounding boxes around damage and classifying it. Because the orthophotos are georeferenced, the output is not just a list of detections but a map of where repairs are needed, which is precisely the information a maintenance program requires to prioritize its limited budget.

Why roofs? In traditional village architecture, the roof is both the most vulnerable element and the most informative one. Tile slippage, ridge failure, and structural sagging are visible from above in ways that foundation problems are not, and roof damage is often the first sign that a building has passed from routine aging into active decline. A system that can detect roof damage across an entire settlement gives conservators an early-warning capability that manual inspection, constrained by narrow lanes, dense building clusters, and limited personnel, has never been able to provide at village scale.

The implications reach well beyond Guangdong. The authors frame their work as a response to a global problem: rural cultural heritage sites worldwide are losing maintenance attention as populations shift to cities, and the buildings most at risk are often those in the most remote locations. A workflow that requires only a consumer-grade drone, a photogrammetry pipeline, and a trained detection model could be adapted to other regions and other architectural traditions, provided that local training data is annotated to teach the model the specific damage patterns of each building type. The open-access publication of the study, under a Creative Commons Attribution 4.0 license, lowers the barrier for other teams to build on the approach.

There is also a quieter, more poignant dimension to the work. Every orthophoto and three-dimensional model in the digital archive is a record of a settlement as it stood on the day of the flight. For villages that ultimately cannot be saved, those archives become the definitive documentation of what was lost, available to researchers, descendants, and future reconstruction efforts. For villages that can be saved, the same data directs repairs to the buildings that need them most. Either way, the combination of 12,762 drone images, 41,213 annotated training instances, a detection model running at 5.77 milliseconds per frame, and a mean average precision of 0.832 represents a genuine shift in how the world’s most endangered rural heritage can be watched over, one flight and one frame at a time.

Subject of Research: UAV photogrammetry and deep learning for digital preservation and automated roof damage detection in Guangdong traditional villages

Article Title: Digital preservation and automated roof damage detection for traditional village heritage in Guangdong using UAV photogrammetry and deep learning

Article References: Zhu, Z., Hsieh, C.-M., Liu, X., Li, X., Zhong, Z., & Chen, Z. (2026). Digital preservation and automated roof damage detection for traditional village heritage in Guangdong using UAV photogrammetry and deep learning. npj Heritage Science. https://doi.org/10.1038/s40494-026-03055-9

Image Credits: AI Generated

DOI: 10.1038/s40494-026-03055-9

Keywords: UAV photogrammetry, deep learning, YOLOv8, roof damage detection, traditional villages, cultural heritage, Guangdong, digital preservation, orthophoto maps, heritage conservation, object detection, rural settlements

The research was conducted during the 2023 Kharif season at an experimental field operated by Praanadhaara Organised Agro Forestry Private Limited in Bapatla District, Andhra Pradesh, a tropical region where monsoon rainfall makes rice the dominant crop. The team, drawn from institutions including Acharya N. G. Ranga Agricultural University, the ICAR-Indian Institute of Rice Research, and South Dakota State University, cultivated three plots of roughly two acres each: broadcasted direct seeded rice, wet direct seeded rice, and transplanted rice. Each method leaves a distinct signature in the field, from the scatter pattern of broadcast seed to the flooded, orderly rows of transplanted seedlings, but capturing that signature requires imagery far sharper than what orbiting satellites can provide.

Using a DJI Mavic 2 Pro drone equipped with a 20-megapixel Hasselblad camera, the researchers flew automated waypoint missions on three dates: October 15, 2023, October 27, 2023, and December 1, 2023. Flights were held to a constant altitude of 40 meters, with 80 percent forward and 75 percent side image overlap, under clear skies and gentle winds between 3 and 4 meters per second. Each mission produced roughly 425 raw red-green-blue images at a ground sampling distance of 1.1 centimeters per pixel. That resolution is extraordinary by remote sensing standards: previous efforts to map direct seeded rice relied on satellite sensors with resolutions between 10 and 250 meters, which blur out the fine structural cues that separate one planting method from another.

After the raw images were orthorectified into georeferenced mosaics, the team overlaid GPS-derived plot boundaries using QGIS software and cut the mosaics into 400-by-400-pixel patches. The resulting dataset contained 1,966 labelled patches: 672 of broadcasted direct seeded rice, 612 of transplanted rice, and 682 of wet direct seeded rice. Because all three systems were managed under identical agronomic conditions and sowing schedules were controlled, any differences the models detected could be attributed to the establishment method itself rather than to confounding variation in crop age or management.

The classification pipeline paired a convolutional neural network with a conventional machine learning classifier. ResNet-50, a 50-layer deep residual architecture, served as the feature extractor, converting each image patch into a 2,048-dimensional vector that encodes spatial and textural information. The authors chose ResNet-50 deliberately: its transfer learning capability and moderate complexity suit relatively small drone datasets, where deeper networks would risk overfitting. Those feature vectors were then fed to six different classifiers: k-nearest neighbor, support vector machine, decision tree, random forest, Naïve Bayes, and a neural network, producing six hybrid models. Two additional end-to-end architectures, YOLOv5 and YOLOv8, were trained on the same data for comparison.

Performance was tested rigorously. The dataset was split across nine different training-to-testing ratios, from 10:90 up to 90:10, and evaluated with overall accuracy, precision, recall, F1-score, the kappa coefficient, and mean average precision. The clear winner was the ResNet-50 plus neural network combination, which achieved average overall accuracies of 93.16 percent, 95.81 percent, and 93.31 percent across the three acquisition dates. No other model matched that consistency. The next best, ResNet-50 plus support vector machine, reached 92.07, 95.24, and 92.09 percent, while the weakest performers, Naïve Bayes and decision tree hybrids, hovered in the high 70s. The YOLO models, despite their popularity in agricultural computer vision, landed in the middle at roughly 78 to 89 percent.

The winning model also posted the strongest secondary metrics, with precision between 0.93 and 0.97, recall between 0.92 and 0.97, F1-scores from 0.92 to 0.97, and kappa coefficients of 0.89 to 0.96. The kappa coefficient is particularly telling, because it corrects for agreement that could occur by chance, meaning the model was genuinely discriminating between classes rather than exploiting class imbalance. Notably, the peak performance came at the second acquisition date, when direct seeded rice was mid-establishment, suggesting the features that distinguish planting methods become most legible partway through the season.

The authors are candid about the study’s limitations. Because transplanted rice was not put into the ground until November 3, the first two drone flights captured that treatment only as pre-transplant field preparation, so early-stage classifications partly reflect soil, water, and land preparation differences rather than a mature transplanted canopy. The design also synchronized sowing dates across treatments, whereas real-world farms stagger planting according to rainfall, water availability, and labor. At regional scales, such phenological variation could complicate classification, and the authors recommend validating the framework across multiple seasons, planting dates, and regions before operational deployment.

Even so, the implications are substantial. Knowing which establishment method dominates a given landscape matters for precision agriculture: transplanted rice, wet direct seeded rice, and broadcast rice differ sharply in water consumption, labor demands, methane emissions, and input requirements. Accurate, fine-grained maps of establishment methods could sharpen area estimates, guide site-specific management, and support policies promoting water-saving direct seeding. The authors point toward future upgrades, including multispectral and hyperspectral sensors to capture plant physiology, transformer and self-supervised architectures that need less labelled data, explainable AI for transparency, and lightweight models that run in real time aboard the drone itself. For now, the study stands as a compelling proof of concept: an off-the-shelf drone, a pretrained network, and a well-chosen classifier can read the hidden history of how a rice field was planted, written in centimeter-scale patterns of leaves, soil, and water.

Subject of Research: Deep learning classification of rice establishment methods using high-resolution drone imagery

Article Title: Deep learning-based classification of wet direct seeded rice, broadcasted direct seeded rice and transplanted rice using drone imagery for precision agriculture

Article References: Gubbala, A. L., Rathod, S., Dasyam, R., Rapolu, M. K., Dasari, A. K., Kurra, P., Madireddy, H. R., Shashishekhar, P. R., Mural, R. V., Kumar, A., & Sundaram, R. M. (2026). Deep learning-based classification of wet direct seeded rice, broadcasted direct seeded rice and transplanted rice using drone imagery for precision agriculture. Frontiers in Remote Sensing, 7, Article 1698781. https://doi.org/10.3389/frsen.2026.1698781

Image Credits: AI Generated

DOI: 10.3389/frsen.2026.1698781

Keywords: deep learning, drone imagery, precision agriculture, remote sensing, ResNet-50, YOLOv8, rice, direct seeded rice, transplanted rice, UAV, image classification, crop mapping

The problem the researchers set out to solve is fundamentally one of scarcity and visual difficulty. Deep learning detectors need large, well-annotated datasets, but collecting and labeling underwater imagery is labor-intensive and expensive. Existing resources such as Trash-ICRA19, with 5,720 images from marine remotely operated vehicle surveys, TrashCan 1.0 with 7,212 segmented images, UNO, CleanSea, and DeepPlastic all focus on marine, coastal, or deep-sea settings. Those environments differ substantially from shallow freshwater ponds in illumination, water clarity, object appearance, and background complexity, so models trained on them transfer poorly to pond conditions. The team also noted that prior freshwater-focused systems, such as camera-based floating plastic detectors and the YOLO-EcoFocus framework for inland waterways, target debris at or near the water surface rather than fully submerged waste on a pond floor.

PondDebris comprises 1,526 images and 3,224 bounding box annotations spread across two ecologically meaningful categories: biodegradable and non-biodegradable waste. Rather than simply sorting trash by material, the taxonomy reflects environmental persistence, a distinction that matters for automated waste management because biodegradable items such as fruit peels, vegetable matter, leaves, uncoated paper, and untreated wood will break down naturally within weeks to a few months, while plastic bottles, plastic bags, aluminum cans, polystyrene, and synthetic packaging will not. Under the annotators’ operational rule, treated or painted wood and coated paper counted as non-biodegradable because their surface treatments delay natural decomposition. The final tally included 1,537 biodegradable and 1,687 non-biodegradable instances.

The imagery was collected in a controlled laboratory aquarium designed to emulate the visual character of a shallow freshwater pond, complete with a sandy substrate, variable turbidity, changing illumination, water reflections, and partial occlusion of objects. A Crosstour CT8500 action camera recorded 4K UHD video across three sessions, the last under different artificial lighting to broaden visual diversity. Frames were extracted, screened to remove redundant or blurred shots, and resized to 512 by 512 pixels for training. The authors are explicit that this is a simulated proxy rather than an in-situ field dataset, a limitation they flag for future work, but the controlled setup allowed precise manipulation of debris placement and environmental variability while demonstrating that useful underwater benchmarks can be built with consumer-grade hardware.

Annotation quality received careful statistical treatment. Three annotators labeled the images in Roboflow, reviewed every annotation twice, and submitted a subset for independent verification by a faculty supervisor. On a stratified subsample re-annotated independently, class-level agreement measured by Cohen’s kappa reached 0.83, and bounding-box localization agreement, expressed as mean intersection-over-union, reached 0.87, both indicating substantial consistency. The dataset was split into 1,206 training, 183 validation, and 137 test images, and because the collection is modest in scale, the team supplemented the single split with five-fold cross-validation to guard against sampling variability. The dataset is released under a CC BY 4.0 license.

On the modeling side, UWAD-Net builds on YOLOv8s, the small variant of a single-stage detector that pairs a cross-stage-partial backbone with an anchor-free, decoupled head that separates classification from localization. YOLOv8s was chosen because it offered the best balance of accuracy and efficiency among six YOLO-family baselines evaluated, running at roughly 182 frames per second on an NVIDIA Tesla T4 GPU with about 11.17 million parameters. The key innovation is the integration of the Convolutional Block Attention Module, or CBAM, into the YOLOv8s architecture. CBAM applies channel attention and spatial attention in sequence: channel attention uses average and max pooling with a shared multilayer perceptron to learn which feature channels carry discriminative information, while spatial attention generates a two-dimensional map through a convolution that highlights informative regions and suppresses background clutter.

The placement of these attention modules was itself a design decision tested empirically. CBAM blocks were inserted after the P3, P4, and P5 feature pyramid levels produced by the YOLOv8s neck, immediately before the detection head, so that each scale receives refinement tuned to small, medium, and large objects respectively. In the ablation study, channel attention alone outperformed spatial attention alone, and inserting full CBAM at all three scales yielded the best overall configuration. The intuition is that underwater imagery suffers from wavelength-dependent light absorption, scattering, and turbidity, so channel attention can emphasize features that survive color degradation while spatial attention suppresses diffuse responses from suspended particles and substrate textures, sharpening the network’s focus on debris boundaries.

Trained for up to 150 epochs with the Adam optimizer, cosine learning-rate decay, and standard augmentations such as mosaic, horizontal flips, and color jitter, UWAD-Net achieved a mean average precision at IoU 0.50 of 0.7294, an mAP50-95 of 0.4517, precision of 0.8206, and recall of 0.6421 on the PondDebris test set, with five-fold cross-validation standard deviations of only a few thousandths, indicating stable performance across data partitions. The most consistent advantage over the YOLOv8s baseline was precision, which rose by 7.1 percentage points, meaning fewer false-positive detections, though the authors are candid that this came with a modest reduction in recall and slightly lower mAP, a precision-oriented trade-off rather than a uniform improvement. Against non-YOLO baselines, UWAD-Net outperformed Faster R-CNN with a ResNet-50-FPN backbone and the transformer-based RT-DETR on the same protocol, and the added attention machinery cost only a marginal increase in computation, keeping the model practical for resource-constrained edge deployment.

To probe generalization, the team fine-tuned a PondDebris-pretrained UWAD-Net on an external Underwater Debris Detection Dataset of 5,125 images spanning 15 litter categories, mapped onto the same two-class taxonomy. The transferred model reached an mAP50 of 0.6629 and precision of 0.7738 on the external test split, confirming that the learned underwater debris representations remain useful in a previously unseen environment, although recall dropped to 0.5724. The degradation reflects domain shift: the external data were captured at roughly 4 to 5 meters depth under medium turbidity, with irregularly shaped items such as wood and leaves proving harder to distinguish from the substrate than well-defined objects like bottles. The authors conclude that improving robustness will require larger, class-balanced datasets and domain-adaptation strategies, and that field-collected pond imagery is the natural next step for validating the framework in real freshwater ecosystems.

Subject of Research: Attention-enhanced deep learning for detecting biodegradable and non-biodegradable waste in freshwater pond environments

Article Title: UWAD-Net: attention-enhanced underwater waste detection in freshwater pond environments

Article References: S, B. B., Rathnasekaran, S. R., Patnaik, R. A., Anbalagan, S., & M, V. (2026). UWAD-Net: attention-enhanced underwater waste detection in freshwater pond environments. Frontiers in Water, 8, Article 1923979. https://doi.org/10.3389/frwa.2026.1923979

Image Credits: AI Generated

DOI: 10.3389/frwa.2026.1923979

Keywords: underwater waste detection, freshwater ponds, UWAD-Net, PondDebris dataset, CBAM attention, YOLOv8, object detection, computer vision, environmental monitoring, plastic pollution, transfer learning, deep learning

The stakes of this challenge are far higher than the small size of the vessels involved might suggest. Kawasaki disease, first described by Dr. Tomisaku Kawasaki in 1967, is an acute inflammatory condition of blood vessels that strikes primarily in early childhood and has a peculiar taste for the coronary arteries, the vessels that feed the heart muscle itself. When the disease goes untreated, the inflamed artery walls can weaken and balloon into dilations or aneurysms, which in turn set the stage for blood clots, narrowing of the vessels, and even heart attack in young patients. Because these coronary complications develop mostly during the acute phase of illness, everything depends on spotting subtle enlargements early, and then tracking vessel dimensions meticulously over years of follow-up to gauge each child’s future cardiovascular risk.

Echocardiography, the ultrasound-based imaging technique at the heart of this study, is the workhorse tool for that surveillance. It is non-invasive, reproducible, fast, and capable of showing the beating heart in real time, which makes it well suited for examining children. But it comes with a catch. Children’s coronary arteries are inherently tiny, their caliber changes as they grow, and on ultrasound they often appear with low contrast and blurred boundaries. Manual measurement of such structures leans heavily on the operator’s experience, and different observers frequently arrive at noticeably different numbers for the same vessel. That variability is problematic enough in major hospitals, but the researchers behind the new work point out that in many parts of China, and by extension in many resource-limited settings worldwide, pediatric echocardiography expertise is scarce, and achieving consistent coronary measurements is a major clinical hurdle for less experienced physicians.

To build their system, the team assembled a substantial dataset from routine clinical practice. They retrospectively collected 2,314 echocardiographic examinations, encompassing 8,766 coronary artery images, from children with suspected or diagnosed Kawasaki disease who were scanned at their hospital between January 2021 and January 2024. All the exams were performed with Philips EPIQ7 or EPIQ5 ultrasound systems while the children were asleep or calm, following the standard imaging views laid out in the American Heart Association’s guidelines. From this pool, the researchers curated 3,000 high-quality images and had two physicians with more than five years of ultrasound experience annotate four key coronary branches: the right coronary artery, the left main coronary artery, the left anterior descending artery, and the left circumflex artery. A senior pediatric echocardiographer then reviewed and corrected every annotation before the dataset was split into training, validation, and test sets in a 7:2:1 ratio.

The technical backbone of the model is YOLOv8, an open-source object detection architecture known for speed and precision. Its design splits the workload across three cooperating components. The backbone, built from so-called C2f modules, extracts visual features in alternating compression and expansion stages, using convolutions to shrink the feature maps and reduce computational load, then up-sampling to restore detail and integrate what has been learned. The neck fuses feature maps from multiple scales, combining specialized modules that let the network reason about both coarse context and fine structure, an essential capability when the targets are small, low-contrast blood vessels. The head is decoupled into separate detection and classification branches, which reduces interference between locating objects and labeling them. The researchers further tuned the pipeline for pediatric ultrasound by optimizing input resolution and data augmentation, randomly flipping, rotating, scaling, and adjusting the brightness and contrast of training images. Training ran for up to 100 epochs on an NVIDIA RTX 3090 graphics card, and a Task-Aligned Assigner along with Distribution Focal Loss helped the model cope with hard-to-detect vessels and class imbalance during learning.

Turning detection into an actual diameter measurement required an additional layer of geometric engineering. After the model locates a vessel, the software first normalizes scale, converting pixel distances into physical lengths using the calibration parameters stored in each image. It then extracts the vessel’s contour, applies a convex hull algorithm to select evenly distributed reference points, and constructs an auxiliary circle centered on each point with a radius set to half the bounding box diagonal. The intersections between that circle and the vessel outline become anchors from which perpendicular lines are drawn into the vessel interior; the lengths of these perpendiculars, averaged over multiple trials, yield the vessel diameter. The pipeline is a clever marriage of modern neural detection with classical geometry, designed specifically to handle the narrow, tortuous anatomy of pediatric coronary arteries.

When the finished model was unleashed on the test set, its performance was striking. For the right coronary artery, it achieved a recall of 96.4 percent, a precision of 94.0 percent, and a mean average precision of 99.2 percent. The left main coronary artery followed closely at 93.4 percent recall, 90.1 percent precision, and 97.0 percent mean average precision, while the left anterior descending artery scored 90.1 and 85.2 percent on recall and precision with a mean average precision of 94.1 percent. The left circumflex artery proved hardest of all, with all metrics lower than for the other branches, likely reflecting its smaller diameter, deeper anatomical position, and poorer visibility on ultrasound. Across the validation set, the model’s classifications agreed with expert physicians at a Kappa statistic of 0.904, a level conventionally interpreted as almost perfect agreement.

The comparisons with human observers were particularly revealing. Expert physicians spotted the right coronary artery and left main coronary artery in 98.4 and 98.8 percent of exams respectively, while general physicians managed 92.7 and 93.1 percent, and the gap widened dramatically for the more elusive branches: experts found the left anterior descending and left circumflex arteries in 84.9 and 57.7 percent of cases, versus just 56.0 and 37.3 percent for general physicians. Age also mattered, with recognition of the right coronary and left main arteries significantly lower in infants than in toddlers and older children, plausibly because babies’ faster breathing and heart rates complicate image acquisition. For measurements, the AI’s agreement with experts was excellent across all branches, with intraclass correlation coefficients all above 0.85 and the narrowest limits of agreement in Bland-Altman analysis, especially for the right coronary and left anterior descending arteries. Agreement between general physicians and either experts or the AI was noticeably looser, with the left circumflex artery again the weakest link for everyone. On speed, the machine was in a different league entirely: less than two seconds per measurement, compared with 16.3 seconds for experts and 18.9 seconds for general physicians using manual calipers.

The authors are candid about the study’s limitations. The model was trained and tested on static images rather than dynamic video, so it misses temporal information about moving vessels; it came from a single center using a single ultrasound brand, which may limit generalizability; the mid-segment of the right coronary artery was excluded due to sparse samples; and the bounding-box-plus-geometry approach can introduce errors in irregular vessels or poor-quality images. Future work, they say, will incorporate echocardiographic video with recurrent or transformer-based temporal modeling, refine the architecture for the difficult left anterior descending and left circumflex branches, and expand to multicenter, multi-device datasets followed by prospective clinical validation. Even so, the implications are considerable. A system that brings expert-level coronary assessment to clinics where such expertise is rare could standardize the diagnosis and follow-up of Kawasaki disease, support quality control in busy imaging departments, and help ensure that the children most at risk of lifelong cardiac damage are caught in time. The researchers caution that challenging vessels like the left circumflex artery still demand expert oversight, positioning the AI not as a replacement for skilled sonographers but as a powerful, tireless assistant, one capable of measuring a child’s coronary arteries faster than a heartbeat.

Subject of Research: A deep learning model for automated identification and measurement of coronary arteries in pediatric echocardiography for Kawasaki disease assessment

Article Title: Deep learning for coronary arteries identification and measurement in pediatric echocardiography

Article References: Lai, L.-P., Li, Z.-L., Liu, Q.-F., Jiang, J.-Y., Wang, L.-Q., He, M.-Y., Lin, M., Li, X.-L., & Li, S.-P. (2026). Deep learning for coronary arteries identification and measurement in pediatric echocardiography. Pediatric Research. https://doi.org/10.1038/s41390-026-05472-9

Image Credits: AI Generated

DOI: 10.1038/s41390-026-05472-9

Keywords: deep learning, Kawasaki disease, pediatric echocardiography, coronary arteries, YOLOv8, artificial intelligence, medical imaging, ultrasound, coronary artery aneurysm, object detection, cardiology, clinical AI

The research, conducted by Li Guo, Wanchen Zhang, Nan Jia, and Caixuan Wang of Information Engineering University in Zhengzhou, China, addresses a long-standing bottleneck in event-oriented information extraction. Textual reports contain rich narrative detail but no explicit coordinates. Web images show visual scenes but no machine-readable references to place. Scanned historical maps encode precise geographic extents, yet their coordinate annotations are locked inside pixels that standard text-processing pipelines cannot touch. The team’s framework chains together textual entity, relation, and event extraction; visual object detection; scanned-map coordinate recognition; geographic entity grounding; cross-modal entity fusion; provenance-aware storage; and graph visualization, so that each modality contributes its own strengths to a unified whole.

One of the most technically demanding pieces of the pipeline is the recovery of geographic information from scanned maps. Old or digitized maps often carry coordinate annotations along their margins, but these numeric labels sit amid decorative cartographic elements, gridlines, and noise. The researchers designed a cascaded procedure that pairs two established open-source tools: YOLOv5, a convolutional neural network for object detection, and PaddleOCR, a text recognition engine. In the first stage, YOLOv5 detects and crops the coordinate annotation regions directly from the scanned image. Only then does PaddleOCR read the numbers within those focused crops, rather than attempting to transcribe the entire cluttered map sheet. The recognized coordinates are subsequently used to reconstruct the geographic extent of the map, effectively giving the digital archive a machine-readable bounding box in the real world.

The quantitative gains from this cascade are striking. On a dataset of 287 scanned maps, YOLOv5 achieved a region-detection F1-score of 0.977, indicating that the detection stage rarely misses a coordinate block. In a paired evaluation across 278 maps using an out-of-fold protocol, the proportion of maps whose coordinates were completely and correctly recognized rose from 64.03 percent with direct PaddleOCR alone to 87.05 percent with the cascaded procedure. Because the same maps were processed both ways, the team could apply an exact McNemar test to the paired outcomes; the improvement was statistically significant, with p less than 0.001. For anyone who has watched OCR systems flounder on dense cartographic layouts, the lesson is clear: detecting where to look before reading what is there pays substantial dividends.

Textual processing forms the narrative backbone of the framework. Entity, relation, and event extraction models applied to conflict-event reports achieved F1-scores between 0.76 and 0.84 across the evaluated tasks, meaning the system can pull out actors, actions, and locations from prose with reasonably high precision and recall. Extracted place entities are then geocoded against GeoNames, the widely used open geographic database, converting place names into latitude-longitude anchors. The crucial connection to maps happens next: textual place entities are linked to relevant scanned maps through point-polygon spatial intersection. If a geocoded location falls inside the reconstructed extent of a map, the report and the map become associated in the graph, allowing an analyst who reads about a village attack to jump straight to the cartographic coverage of that area.

Visual evidence enters the pipeline through modern object detection. The team evaluated YOLOv8 on three categories of web imagery and reported mean average precision at an intersection-over-union threshold of 0.5 — mAP@0.5 — of 0.935 for aircraft, 0.890 for ships, and 0.922 for general web-image detection. These are demanding visual recognition problems: aircraft appear at varying scales and angles in satellite or aerial imagery, and ships occupy few pixels against reflective water. The high scores suggest that the detection components are mature enough to feed reliable visual nodes into the graph, each image contributing detected objects that can be associated with events, times, and places.

Perhaps the most consequential step is cross-modal entity fusion, the process of deciding that a place name in a report, a detected object in a photograph, and a coordinate range on a map all refer to the same real-world thing. Naive matching fails badly in conflict contexts, where place names repeat across regions and abbreviations abound. The framework therefore applies a constrained fusion method, and the evaluation design matters as much as the method itself: the team used an event- and source-document-grouped protocol, which prevents information leakage between training and test cases that share events or documents. Under this strict protocol, the complete constrained cross-modal entity-fusion method achieved an F1-score of 0.884, a strong result for a task that spans modalities as different as narrative text and raster graphics.

Provenance-aware storage is the quiet innovation that makes the resulting graph trustworthy. Every extracted fact, whether it came from a sentence, a detected object, or a recognized coordinate label, carries a record of its source document and extraction method. This matters in conflict analysis, where contradictory reports are common and the credibility of a claim often depends on where it originated. By storing provenance alongside the fused entities and events, the graph does not flatten disagreement into a single answer; it preserves the evidentiary chain so that analysts can audit why the system linked a photograph to a place or an event to a date. Graph visualization then renders the structure interactively, letting users traverse from an event node to its actors, locations, timestamps, source texts, images, and associated maps.

To demonstrate practical applicability, the authors include a conflict-event case study showing the full pipeline in action: reports are mined for entities and events, web images are scanned for relevant objects, scanned maps are georeferenced through the cascaded OCR procedure, and the resulting graph ties narrative, visual, and cartographic evidence together for a real analysis scenario. The case study positions the framework as more than a collection of benchmark scores; it is an end-to-end system aimed at the daily workflow of analysts who currently juggle disconnected tools.

The broader significance of this work lies in how it treats multimodality as a first-class requirement rather than an afterthought. Knowledge graphs have become a standard substrate for intelligence and crisis informatics, but most operational systems still ingest text primarily, with images and maps handled as attachments rather than integrated knowledge. By demonstrating statistically validated improvements on the hardest step — reading coordinates off scanned maps — and a rigorous, leakage-resistant evaluation of cross-modal fusion, the Zhengzhou team has provided both a reusable method and an honest yardstick for future systems. As archives of scanned maps and multimedia reporting continue to grow, frameworks of this kind may determine how quickly raw, scattered evidence becomes structured, queryable knowledge about the world’s most urgent events.

Subject of Research: Multimodal spatiotemporal knowledge graph construction for conflict-event analysis

Article Title: Multimodal spatiotemporal knowledge graph construction for conflict-event analysis: an integrated framework with geographic entity-map association

Article References: Guo, L., Zhang, W., Jia, N., & Wang, C. (2026). Multimodal spatiotemporal knowledge graph construction for conflict-event analysis: an integrated framework with geographic entity-map association. Scientific Reports. https://doi.org/10.1038/s41598-026-73358-1

Image Credits: AI Generated

DOI: 10.1038/s41598-026-73358-1

Keywords: knowledge graph, multimodal, spatiotemporal, conflict events, geographic grounding, entity fusion, OCR, YOLOv5, YOLOv8, PaddleOCR, geocoding, information extraction