The research, authored by Yukun Zhang of the Academy of Fine Arts at Weinan Normal University in China, addresses a fundamental tension in digital heritage protection. Watermarks must be imperceptible, because any visible alteration to dense patterns, delicate local decorations, or stylistic structures compromises both the aesthetic value and the cultural authenticity of the artwork. Yet watermarks must also be robust, because real-world transmission subjects images to platform compression, cropping, filtering, noise enhancement, and re-photography. If only robustness is emphasized, embedding disturbances can damage the fine local structures that make heritage art distinctive. If only invisibility is emphasized, the watermark quickly fails once the image is disseminated. Traditional watermarking schemes, the study argues, have never adequately modeled the full composite distortion distribution that heritage images actually encounter online.

Zhang’s solution begins with an attack simulator called the Hybrid-Adaptive Image Watermark Attack Network, or HAIWAN. Current protection methods typically rely on differentiable attack layers — simplified mathematical stand-ins for real transmission distortions — but these layers fail to capture the nonlinear, composite, and highly concealed nature of actual online interference. HAIWAN is designed as an independent strong-sample generator, deliberately separated from the watermark embedding backbone. During training, it continuously produces difficult attack samples that expand the distortion distribution the protection model must withstand. Crucially, because HAIWAN is excluded from the embedding and extraction processes during inference, it expands the training disturbance distribution without adding any computational cost to the deployed system.

Technically, HAIWAN reorganizes three components into a unified attack-generation path. First, a local stability enhancement unit — the Fast and Precise Block — applies residual feature updates while eliminating batch normalization, preserving the continuity of shallow perturbation features without heavy computational overhead. This architecture strengthens the network’s perception of local texture patterns, watermarks, and subtle texture correlations, which matters enormously for heritage artworks with dense patterns and delicate local ornaments. Second, parallel dilated convolution branches with multiple dilation rates replace continuous downsampling, expanding the receptive field while preserving the image’s main structure and capturing perturbation patterns at different spatial scales simultaneously. Third, constrained decoding and residual reconstruction convert the aggregated features into a deep perturbation residual that is added directly to the watermarked image. The result is an attack sample that retains recognizable structure — a weighted peak signal-to-noise ratio of 35.72 decibels and a structural similarity of 0.9187 — while inflicting maximum damage on watermark recovery.

How effective is this attack? Zhang compared HAIWAN against two baseline attack algorithms, WatermarkAttacker and WEvade, under unified inputs and consistent evaluation procedures. On three representative test images, HAIWAN produced bit error rates ranging from 18.18 to 20.98 percent and normalized correlation values from 75.21 to 77.97 percent — meaning it caused greater watermark extraction errors than either competitor. Notably, the normalized correlation values remained above 0.75, showing that substantial watermark correlation was preserved and the watermark was not completely removed. This is precisely the point: HAIWAN is not built as a weapon to strip watermarks from published images, but as a training adversary that generates difficult samples so the protection model learns to survive them. In cross-method generalization tests, HAIWAN maintained stronger destructive capability across different watermarking techniques, with bit error rates ranging from 0.5014 against a conventional DWT-DCT scheme to 0.3529 against the study’s own protection model — evidence that its multi-scale disturbance modeling interferes with both embedding traces and recovery features across diverse watermark mechanisms.

The second layer of the framework is the HAIWAN-guided Dual-domain Robust Blind Watermarking Protection model, abbreviated HDRB-WP. Rather than perturbing the full spatial domain — which risks irreversible damage to detail regions in texture-sensitive heritage images — the model exploits a dual-domain carrier built from the luminance channel. The luminance component retains principal intensity and contour information even after compression and screen recapture, so the low-frequency subband of the discrete wavelet transform provides a structurally stable embedding region. Within that subband, the discrete cosine transform maps the carrier into frequency-domain coefficients, allowing coefficient-level regulation of embedding modifications. This combination improves resistance to compression and filtering while reducing concentrated modification of the texture-sensitive regions that define heritage art.

Inside HDRB-WP, a Channel Attention Embedding Unit performs joint channel-spatial recalibration to decide where copyright information should be written. Channel attention emphasizes frequency-band information most relevant to copyright embedding, while spatial attention selects positions better suited to carrying hidden perturbations — effectively steering the watermark away from key decorative regions of the artwork. The embedding network then completes copyright writing through residual updates governed by an embedding strength coefficient. On the extraction side, a blind decoding network takes the dual-domain representation as input, uses residual feature enhancement as its core, and recovers the copyright mark under strong perturbation without requiring the original image. Training is driven by a joint optimization objective balancing image fidelity — measured with both pixel-level deviation and the perceptual LPIPS metric — watermark recovery through binary cross-entropy, adversarial constraints from a discriminator that pushes watermarked images to be indistinguishable from originals, and a robustness penalty computed over the full attack distribution.

The performance numbers are striking. The complete HDRB-WP model reduced the watermark bit error rate to 0.0169 and achieved watermark detection accuracy of 98.68 percent, outperforming the invertible-network IWRN scheme, which recorded 0.0284 and 97.26 percent respectively, and the single-encoder dual-decoder SEDD scheme. Under five conventional attack types — JPEG 50 compression, Gaussian noise, 3×3 median filtering, 10 percent cropping, and screen recapture — HDRB-WP controlled bit error rates within 0.0622 to 0.1119 on representative samples, while IWRN reached 0.1212 to 0.2248. The gap was most pronounced under screen recapture conditions, the scenario that most closely mimics real-world heritage image dissemination through display-camera recapture. Ablation experiments confirmed that every module contributes: removing the residual enhancement block degraded recovery accuracy, removing the dual-domain branch hurt both visual quality and recovery, and removing either the channel or spatial recalibration branch of the attention unit measurably weakened performance.

Perhaps the most compelling evidence for the attack-simulation approach comes from the guided training gain analysis. Before HAIWAN-guided training, HDRB-WP’s bit error rates under strong disturbances — JPEG 20, Gaussian noise with σ=30/255, 7×7 median filtering, 25 percent cropping, brightness adjustment of 1.35, and enhanced screen recapture — ranged from 15.48 to 29.87 percent. After attack simulation training, those rates dropped to 9.86 to 20.74 percent, with detection accuracy rising from 73.48–87.17 percent to 81.95–93.12 percent and normalized correlation climbing to as high as 0.9084. The improvements were largest under cropping and screen recapture, the distortions that conventional differentiable attack layers model least adequately. The deployed model remains lightweight: 4.89 million parameters, 29.63 milliseconds for embedding, and 17.38 milliseconds for extraction per image — fast enough for low-latency copyright verification in media-integrated communications.

The framework is not without limits. Under geometric desynchronization attacks — rotation of ±5 degrees, scaling by 0.75 or 1.25, 10 percent translation, and 5 percent perspective displacement — bit error rates ranged from 0.1816 to 0.2847, because HDRB-WP lacks a dedicated geometric registration module; crop and resize augmentation provides only partial tolerance. The study also notes that its reported robustness concerns payload recovery under transmission distortions rather than cryptographic resistance to key recovery, forgery, or collusion, and that intentional attacks still require complementary authentication and access-control mechanisms. Still, the implications for cultural heritage are significant. In exhibitions, the technique can protect the authenticity of intangible cultural heritage art as images circulate digitally; in online dissemination, it can support copyright verification after platform compression, image editing, and screen recapture. Zhang’s future work will pursue lightweight geometric synchronization and conflict-free multi-watermark allocation, enabling multiple institutions to verify collaborative copyright on the same heritage image. For a field where a single glaze pattern or embroidery stitch can embody centuries of cultural identity, an invisible mark that survives the internet’s full gauntlet may prove to be heritage’s most durable signature.

Subject of Research: Digital visual transmission and watermark protection of intangible cultural heritage art images based on the HAIWAN method

Article Title: Digital visual transmission and watermark protection of intangible cultural heritage art images based on the HAIWAN method

Article References: Zhang, Y. (2026). Digital visual transmission and watermark protection of intangible cultural heritage art images based on the HAIWAN method. Discover Artificial Intelligence, 6(1), Article 1537. https://doi.org/10.1007/s44163-026-02523-2

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02523-2

Keywords: Digital, visual, transmission, watermark, protection, intangible, cultural, heritage, images, based, HAIWAN, method

PET/CT, which combines positron emission tomography with computed tomography, is widely used to diagnose, stage and monitor cancer. Conventional PET/CT scanners have a short axial field of view, meaning they image the body section by section, typically requiring five to six bed positions to cover the body from head to thigh. Long-axial field-of-view (LA-FOV) systems capture a much larger portion of the body at once and detect more of the signal emitted by the radiotracer, raising the possibility of much shorter scans without sacrificing image quality.

In the EQUIP study, researchers at the Peter MacCallum Cancer Centre in Melbourne, Australia, evaluated a next-generation 128 cm LA-FOV PET/CT system, the GE HealthCare Omni 128 cm. Twenty-four participants who had received one of five radiotracers underwent a standard-of-care scan on a conventional short-axial field-of-view (SA-FOV) PET/CT system, followed by a ten-minute scan on the LA-FOV system. The researchers then reconstructed images from the LA-FOV scan using just five, three, two and one minute of data.

Two nuclear medicine specialists, unaware of the acquisition times, graded the images on a five-point scale for overall diagnostic quality, sharpness, lesion conspicuity and image noise. Mean image scores were in the good or excellent range at every acquisition time. Average diagnostic quality scores were 5.0 for the ten- and five-minute images, 4.9 for the three- and two-minute images and 4.5 for the one-minute images. By comparison, the conventional scanner scored 3.9 for diagnostic quality, with lower average scores than the LA-FOV system across all assessed categories and acquisition times.

Lead author Dr Cameron Pain, a nuclear medicine physicist at Peter MacCallum, explained the striking nature of the result. A one-minute scan on this system gave images that readers rated more highly than the standard-of-care scans performed every day. Even when the acquisition time was cut by 90 percent, image quality exceeded current standard-of-care methods.

All images acquired for two minutes or longer scored at least four out of five in every category, and all one-minute images scored at least three out of five. Image scores declined as scan times shortened, with the largest decrease at one minute, particularly for image noise. Even so, these one-minute images were closer in quality to the full ten-minute LA-FOV images than to the standard-of-care SA-FOV images. Objective measurements supported the reader findings, with signal-to-noise ratios in healthy liver tissue remaining broadly consistent between the ten-minute and one-minute images for the radiotracers assessed.

Shorter scans have the potential to make PET/CT more comfortable for patients, particularly those who are in pain or find it difficult to lie still, while also reducing image blurring caused by movement. They could also help imaging departments scan more patients each day. Alternatively, the higher sensitivity of LA-FOV systems could be used to reduce the amount of radiotracer administered, lowering the radiation dose patients receive.

Total-body PET gives choices that have not existed before, Dr Pain noted. Its sensitivity can be used to scan patients faster, to lower the radiation dose and to see disease more clearly. Further research will be needed to confirm these findings. Image quality was assessed using reader scores rather than measures of diagnostic accuracy, so larger studies will be important to confirm that very short scans do not miss small or faint lesions and to determine how faster scanning affects workflow and patient experience in routine clinical practice.

The EQUIP findings arrive as long-axial field-of-view total-body PET systems gain regulatory approvals and enter clinical service more broadly. If subsequent studies confirm that one-minute acquisitions preserve diagnostic accuracy, the change could ripple through oncology departments worldwide, easing burdens on patients who struggle with long scans while expanding daily throughput. For a technology long constrained by the physics of counting photons, the ability to compress a scan from ten minutes to one, while exceeding the quality of today’s standard of care, marks a striking demonstration of what next-generation detector sensitivity can deliver.

The physics behind this gain in speed lies in geometric efficiency. In a conventional scanner, a radiotracer signal is only counted when the pair of annihilation photons strikes detectors on opposite sides of a narrow ring, so most photons emitted by the patient escape undetected. A scanner whose detector ring spans roughly the entire torso captures a far larger fraction of that emission, and the sensitivity advantage scales dramatically — often by a factor of several dozen compared with short-axial systems. That surplus of detected signal is what allows reconstruction algorithms to form a usable image from a fraction of the usual counting statistics, and it explains why the one-minute images in the EQUIP study could still compete with, and exceed, conventional standard-of-care acquisitions lasting far longer.

The five radiotracers used in the study also illustrate how versatile such a platform could be. Fluorine-18 FDG remains the workhorse of oncological PET, mapping glucose metabolism across a broad range of cancers. Prostate-specific membrane membrane tracers such as fluorine-18 DCFPyL and gallium-68 PSMA agents target prostate cancer with high specificity, while gallium-68 DOTATATE images neuroendocrine tumours through somatostatin receptors. Iodine-124, a longer-lived positron emitter, serves niche applications where its half-life permits imaging days after injection. Demonstrating consistent image quality across this pharmacological spectrum matters because a scanner’s sensitivity dividend is only clinically meaningful if it holds regardless of which tracer a department deploys.

Reduced acquisition time also interacts with patient physiology in ways that go beyond comfort. PET images are inherently vulnerable to motion: a patient breathing, shifting position, or simply unable to remain still over a five-to-six-bed-position protocol introduces blur that can obscure small lesions. Compressing the acquisition into a single minute narrows the window during which motion can degrade the data, which is particularly relevant for imaging the lungs, liver and other structures that move with respiration. For patients in pain, children, or those with claustrophobia or cognitive impairment, the difference between a tolerable and an intolerable examination can determine whether staging information is obtained at all.

The throughput implications deserve attention as well. Nuclear medicine departments in many health systems face backlogs driven by limited scanner availability and the sequential nature of conventional protocols, in which each patient occupies the table for the full duration of a multi-bed acquisition. If diagnostic-quality images can be obtained in a minute or two, the number of patients scanned per day could rise substantially, or alternatively the spare sensitivity could be traded for lower injected activity — reducing radiation exposure, easing constraints on radiopharmacy production, and potentially making scans feasible for populations, such as pregnant patients or those requiring repeated surveillance imaging, in whom dose is a particular concern.

Objective metrics reinforce the subjective reader scores. Signal-to-noise ratio measured in healthy liver tissue remained broadly consistent between the ten-minute and one-minute reconstructions for the tracers assessed, indicating that the statistical penalty of a shorter acquisition was largely offset by the scanner’s sensitivity rather than by aggressive smoothing or noise-suppression techniques that can mask genuine pathology. This distinction is important, because an image that looks clean but suppresses subtle uptake could mislead clinicians even when quantitative measures appear favourable.

Nevertheless, the study’s design leaves important questions open. Reader scores capture perceived quality — sharpness, conspicuity, noise — but not whether a scan changes a diagnosis, upstages a tumour, or detects a metastasis that would otherwise be missed. Sensitivity analyses powered on clinical endpoints, involving larger and more heterogeneous cohorts, will be needed before one-minute protocols can be adopted with confidence, particularly for small or faint lesions where the one-minute images showed their greatest degradation in noise. How such protocols integrate with CT components of the examination, radiotracer uptake waiting times, and reporting workflows will also shape whether the theoretical throughput gains materialise in practice.

The EQUIP study is registered as a formal clinical trial, reflecting a deliberate effort to move total-body PET from a research curiosity toward routine clinical tool. As long-axial systems receive regulatory clearances and proliferate beyond a handful of pioneering centres, the central question shifts from whether the hardware works to how its extraordinary sensitivity should best be spent — on speed, on dose reduction, or on image clarity. The Melbourne findings suggest these need not be mutually exclusive, and that the trade-offs clinicians have long accepted as inherent to PET may be far more negotiable than previously assumed.

Subject of Research: Total-body PET imaging

Article Title: Next-generation total-body PET scanner can produce diagnostic-quality images in as little as one minute

Article References: Next-generation total-body PET scanner can produce diagnostic-quality images in as little as one minute. Original publication (Bibliographic details unavailable.)

Image Credits: AI Generated

DOI: Not provided

Keywords: Next-generation, total-body, scanner, produce, diagnostic-quality, images, little, minute, scientific research

Segmentation, the pixel-by-pixel outlining of objects of interest within an image, is the bedrock of quantitative pathology. Once a tumor region, a gland, a nucleus, or an inflammatory cell cluster has been precisely delineated in a digitized tissue slide, everything else follows: automated cell counting, measurement of invasive fronts, quantification of stromal composition, and the downstream prediction of diagnosis and prognosis. Yet the traditional route to such measurements has demanded a dedicated deep learning model for each structure of interest, each trained on painstakingly annotated data, or else repeated manual prompting in which a user must click points or drag boxes around every single object in every single image. Neither approach scales gracefully to the enormous structural diversity of human histology.

PathSegmentor takes a different route. Instead of relying on spatial prompts, the model accepts a textual description of the target structure, such as a named anatomical region, histological structure, or object type, and produces the corresponding segmentation mask directly. The language interface means that a researcher or clinician does not need to identify and localize every object before the model can work; a single descriptive phrase suffices. In practice, the team reports, these text prompts substantially reduced the need for object-by-object point or box annotations while remaining robust to natural variation in wording, so that slightly different phrasings of the same request still yield reliable segmentations.

The foundation for this flexibility is PathSeg, an aggregated resource assembled from 21 publicly available pathology image segmentation datasets. Together these contribute 275,200 image–mask–label triples, each pairing a pathology image with its expert ground-truth mask and a categorical label. Crucially, the labels are not a flat, idiosyncratic jumble. The team organized them into a three-level hierarchy that moves from anatomical region, through histological structure, down to the specific object type. This structure allows a single naming scheme to span everything from broad tissue classes to individual epithelial cells, smooth muscle cells, leukocytes, red blood cells, lymphocytes, healthy and cancerous prostate glands, and tumor epithelium, across organs including the breast, colon, lung, and prostate.

Trained on this resource, a single PathSegmentor model achieved the highest overall segmentation performance across 16 internal test datasets, encompassing roughly 45,000 evaluated image–mask pairs. The comparisons were rigorous and broad: the authors benchmarked against task-specific state-of-the-art architectures such as nnU-Net and DeepLabV3+, against specialized pathology adaptations like SAM-Path, and against spatial-prompted segmentation foundation models including MedSAM and SAM-Med2D, as well as the multi-modal biomedical model BiomedParse. Statistical comparisons used one-sided Student’s t-tests with Holm-adjusted p-values and bootstrap-estimated 95 percent confidence intervals, and the model’s advantage was consistent across the hierarchy of anatomical regions, histological structures, and object types, with particularly notable gains on intricate, morphologically complex objects where competing approaches struggled.

Internal performance alone proves little in medical imaging, where domain shift, differences in scanners, staining protocols, patient populations, and disease prevalence, can silently degrade a model that looked flawless in development. The researchers therefore validated PathSegmentor on external public datasets and, importantly, on clinical cohorts contributed by collaborating hospitals, including two clinical datasets whose redistribution is restricted by institutional rules. The model generalized across these external settings and across object categories, suggesting that its text-driven, hierarchy-aware design captures representations of histological structure that transfer beyond the specific datasets it was trained on.

Perhaps the most clinically consequential application demonstrated in the study is interpretability. Deep learning classifiers for breast cancer, including models trained on large whole-slide image collections, are famously opaque: they produce a diagnosis or a biomarker call without a transparent account of which image features drove the decision. The team exploited PathSegmentor’s predicted structures as a lens onto these classification models. By systematically perturbing individual segmented objects and by mapping class activation within the boundaries of each predicted structure, they could attribute a breast cancer classifier’s behavior to specific, nameable histological entities, distinguishing the contributions of tumor regions, healthy glands, and other recognized structures in a way that raw saliency maps cannot.

This object-level explanation capability connects to a broader movement in medical artificial intelligence toward models whose reasoning can be audited by the clinicians who must act on their outputs. Instead of explaining a black box with another approximation of a black box, the approach grounds explanations in concrete anatomical and histological units that pathologists already understand. For cancer diagnosis in particular, where lobular versus ductal morphology, lymphocytic infiltration, and gland architecture all carry distinct clinical meaning, the ability to ask not just what the model predicted but which structures it relied upon could accelerate trust, regulatory review, and clinical adoption of computational pathology tools.

The authors note that their results establish a unified framework for flexible pathology segmentation with potential utility for clinically interpretable image analysis, and the practical infrastructure reflects that ambition. The PathSegmentor source code and analysis scripts are publicly released under an MIT License, with the exact release archived on Zenodo, and PathSeg’s dataset sources and harmonization procedures are documented in the paper’s supplementary materials. Remaining challenges are candidly acknowledged in the authors’ own failure analysis: the model can still struggle with complex or ambiguous boundaries and with densely clustered cells, and the clinical datasets underpinning external validation are available only by request through the contributing hospitals. Even so, the combination of a quarter-million annotated triples, a single model spanning organs and scales, and a natural-language interface represents a significant step toward pathology image analysis in which the question you ask, in words, determines what the machine sees.

The lineage behind PathSegmentor traces back to the Segment Anything Model, the general-purpose vision system introduced in 2023 that demonstrated how a single segmentation model, trained on more than a billion masks, could generalize to unfamiliar images when guided by interactive prompts. That breakthrough quickly inspired medical adaptations, most prominently MedSAM, which showed that the same prompt-driven paradigm could be retuned for radiology, microscopy, and other clinical modalities. What these systems lacked, however, was a way to specify targets semantically. A point or box tells a model where to look, but not what it is looking at, which is precisely the gap that language-prompted approaches such as BiomedParse and PathSegmentor now aim to close in biomedical imaging.

The three-level label hierarchy in PathSeg also reflects a practical truth about pathology annotation: structures in tissue slides are inherently nested. A breast tissue section belongs to an anatomical region, contains histological structures such as glands or stroma, and those structures in turn harbor object types like lymphocytes or tumor epithelium. Encoding this nesting into the training labels gives the model a form of structured supervision that flat label spaces cannot provide, and it mirrors how pathologists themselves are trained to reason about tissue, from organ context down to cellular detail.

The interpretability experiments build on a well-established tension in medical machine learning. Saliency-based explanation techniques, including class activation mapping and randomized input sampling methods, highlight pixels that influence a classifier but produce heatmaps that resist clinical interpretation. Object-level perturbation offers an alternative: by removing or altering a named structure, such as a segmented gland or an infiltrating lymphocyte population, and observing how the classifier’s output changes, the analysis yields attributions expressed in vocabulary that pathologists use daily. This aligns with arguments in the field that high-stakes decisions demand explanations grounded in human-meaningful concepts rather than pixel-level artifacts.

The connection to breast cancer classification is particularly apt given the field’s history. Large molecular studies of breast tumors, including the landmark Cancer Genome Atlas characterization, established that morphological features visible in routine histology correlate with molecular phenotypes, and subsequent work has shown that image-derived features can predict diverse molecular properties of tumors. A segmentation model that reliably delineates the relevant structures provides the missing bridge between such slide-level predictions and the concrete histological evidence behind them.

It is also worth situating the computational approach within the broader methodology of computational pathology. Because whole-slide images are gigapixel-scale and diagnosis often depends on subtle findings scattered across a slide, many classification pipelines rely on multiple-instance learning, in which a slide is treated as a bag of patches with only slide-level labels available. Object-level explanations from a segmentation foundation model could complement these pipelines by revealing which instances within the bag carry diagnostic weight, addressing a long-standing limitation of weakly supervised approaches.

The release of code, an archived version, and documented dataset harmonization procedures follows growing expectations for reproducibility in machine learning for healthcare, allowing independent groups to verify the reported benchmark comparisons and to extend the framework to new organs, stains, and scanning systems as annotated data continue to accumulate.

Subject of Research: Segment anything in pathology images with natural language

Article Title: Segment anything in pathology images with natural language

Article References: Chen, Z., Hou, J., Lin, L., Wang, Y., Bie, Y., Wang, X., Zhou, Y., Li, D., Tan, H., Liang, L., Chan, R. C. K., & Chen, H. (2026). Segment anything in pathology images with natural language. Nature Computational Science. https://doi.org/10.1038/s43588-026-01042-5

Image Credits: AI Generated

DOI: 10.1038/s43588-026-01042-5

Keywords: Segment, anything, pathology, images, natural, language, scientific research