Researchers at the University of Liège have now confronted this question head-on in a review published in Trends in Molecular Medicine, arguing that the field’s long search for a single decisive biomarker may have been looking for the wrong kind of answer. Michael Herfs, an oncologist at the GIGA Institute in Liège, along with colleagues Diane Bruyère and Romane Geuzaine, contends that no viral, immune or molecular marker identified to date can reliably separate self-resolving infections from those destined to progress. Their analysis suggests the problem is not a shortage of candidate markers but a fundamental misunderstanding of what kind of biological system HPV-driven carcinogenesis actually is.

The core of the difficulty is that HPV-associated cancer development is not a linear process with a single molecular switch. It is a dynamic, multi-layered phenomenon in which the outcome depends on the interplay of viral characteristics, the host’s immune response, genetic and epigenetic alterations, the local microbiota, hormonal influences and broader environmental factors. On top of these interacting variables sit stochastic events, most notably the accumulation of mutations driven by virus-induced genomic instability. Two patients infected with the same viral type can follow radically different trajectories because the determinants of progression are distributed across biological scales rather than concentrated in one measurable signal.

This complexity helps explain why decades of biomarker research have produced incremental improvements rather than a decisive breakthrough. Viral load measurements, E6 and E7 oncogene expression, methylation patterns, immune cell profiling and a host of other candidate indicators have each shown promise in isolated studies, yet none has achieved the predictive accuracy needed to transform clinical decision-making. Herfs crystallizes the field’s predicament in a striking observation: perhaps the main challenge is no longer to discover a new biomarker, but to understand how all the biological mechanisms interact to determine the outcome of an HPV infection. The framing represents a conceptual shift from reductionist marker-hunting toward systems-level interpretation.

It is precisely this shift that recent technological developments have made feasible. High-throughput molecular analysis platforms can now generate vast datasets capturing gene expression, epigenetic modification, viral integration sites and immune signatures simultaneously, while artificial intelligence and machine learning approaches offer the computational tools to integrate these heterogeneous data streams into coherent predictive models. Rather than asking whether a single marker predicts progression, integrative models can weigh dozens or hundreds of features at once, potentially capturing the multi-factorial nature of carcinogenesis that has defeated simpler approaches.

The Liège team sees particular promise in using these models to distinguish between the possible trajectories of an infection or a low-grade precancerous lesion: spontaneous regression, persistence without progression, advancement to a more serious lesion, or evolution toward invasive cancer. Bruyère, a researcher at the Laboratory of Experimental Pathology, explains that artificial intelligence and integrative models could help clarify the diversity of biological trajectories observed in patients and ultimately enable a more personalised form of risk assessment. In practical terms, this would mean moving beyond the binary categorizations of current screening toward a graduated risk profile that reflects the actual biology of each individual infection.

The clinical stakes of such a capability are considerable. Current management of HPV-positive screening results and low-grade lesions relies on HPV testing, cytology and histopathological analysis, supplemented by monitoring protocols that are only weakly personalised. This one-size-fits-all approach generates a well-documented burden of unnecessary colposcopies, repeat biopsies and anxiety-inducing follow-up for patients whose infections would have resolved regardless, while simultaneously risking under-surveillance in the minority whose lesions are genuinely dangerous. A validated predictive framework could tailor follow-up intensity to actual risk, reducing overtreatment and its associated costs and harms while concentrating clinical attention where it is most needed.

However, the researchers are careful to temper expectations with an acknowledgment of the current gap between promise and practice. These new models have not yet been sufficiently validated for routine clinical use, and the path from an encouraging computational result to a deployable diagnostic involves rigorous external validation, standardization of input data, regulatory approval and demonstration of clinical utility across diverse populations. The history of biomarker research in oncology offers cautionary examples of tools that performed impressively in retrospective datasets but faltered in prospective clinical testing. The Liège authors position integrative AI approaches as the most promising avenue for improving risk stratification, not as a finished solution.

Their broader conclusion reframes the research agenda for the field. Rather than continuing to pursue the elusive single indicator that will cleanly predict progression, the future lies in a better overall understanding of the biology of HPV infections as integrated biological systems. This perspective aligns with wider trends in cancer research, where multi-omics profiling and machine learning are increasingly applied to problems whose complexity resists traditional reductionist methods. For HPV specifically, the approach could eventually unify viral genomics, host immunology, epigenetics and microbiome data into models that reflect the true determinants of infection outcome.

If that vision matures, the practical consequences could extend well beyond cervical cancer prevention. Persistent oncogenic HPV infection is also implicated in a rising proportion of oropharyngeal cancers and in anogenital malignancies, and improved risk prediction would benefit screening and surveillance strategies across all of these sites. For now, patients and clinicians must continue working with the imperfect tools of cytology, HPV genotyping and histopathology, but the Liège analysis offers a clear-eyed map of where the field must go next: away from the search for a molecular oracle and toward a genuinely integrative understanding of how viral and host biology combine to decide, infection by infection, whether the outcome is clearance or cancer.

Subject of Research: Prediction of progression risk in oncogenic human papillomavirus infection using integrative biomarkers and artificial intelligence

Article Title: Human papillomavirus: Can we finally predict which infections will progress to cancer?

Article References: Human papillomavirus: Can we finally predict which infections will progress to cancer?. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: HPV, human papillomavirus, cancer risk prediction, biomarkers, artificial intelligence, cervical cancer, carcinogenesis, risk stratification, molecular medicine, precancerous lesions, University of Liège, personalised medicine

The clinical stakes are considerable. Gastric cancer remains a leading cause of cancer death worldwide, and its prognosis depends overwhelmingly on the stage at which it is detected. Early gastric cancer, confined to the innermost layers of the stomach wall, can often be treated endoscopically with organ preservation and excellent survival, whereas advanced disease carries a far grimmer outlook. That is why screening and opportunistic upper endoscopy place such emphasis on meticulous inspection. Professional bodies such as the European Society of Gastrointestinal Endoscopy and the Japan Gastroenterological Endoscopy Society have long recognized that incomplete visualization is a major source of missed lesions, and they have promoted systematic inspection protocols that divide the stomach into standard anatomical regions. What has been missing is an automated, quantitative measure of exposure that does not depend on an observer’s subjective judgment of whether a photograph or video frame shows a well-distended, clean, fully visible patch of mucosa.

The research team, led by investigators at Renmin Hospital of Wuhan University together with collaborator Cesare Hassan of Humanitas University in Milan, set out to close that gap. Their approach rests on a deep convolutional neural network trained to perform semantic segmentation of endoscopic images, meaning the model learns to classify every pixel in a frame as belonging to visible gastric mucosa or not. Before training and validating the algorithm, a panel of expert endoscopists defined a reference standard covering 19 anatomical sites of the stomach, spanning the greater and lesser curvatures and the anterior and posterior walls of the gastric body and antrum, along with the cardia, fundus, and angularis. For each of these sites, the experts established a threshold of mucosal exposure that constitutes adequate visualization. When the algorithm’s measured exposure area for a given site exceeds that threshold, the site is classified as adequately exposed; when it falls short, the site is flagged as inadequately exposed.

Technical validation of the segmentation model was carried out using standard metrics for image segmentation performance, including the mean intersection over union, a measure of how closely the algorithm’s pixel-level delineation of mucosa matches expert annotation. But the true test of the OGMEA concept came from clinical data of unusual scale and rigor. Rather than relying on retrospectively collected endoscopy images, the team embedded their analysis within a randomized controlled trial involving 25,320 participants recruited across 24 centers. This multi-center dataset, gathered under a protocol approved by the ethics committee of Renmin Hospital of Wuhan University and the ethics boards of every participating site, with written informed consent from all participants, provided a large, prospectively collected, and heterogeneous body of examinations against which the association between mucosal exposure and lesion detection could be tested.

The results of that analysis are striking. When the researchers classified assessed anatomical sites by the algorithm’s verdict, they found that only 46.36 percent of sites were adequately exposed, while a slight majority, 53.64 percent, fell below the OGMEA threshold. In other words, even in the setting of a controlled trial, more than half of the stomach regions examined were not visualized to the standard defined by the expert panel. This single number is likely to provoke discussion in endoscopy units around the world, because it quantifies, for the first time at this scale, how much room for improvement exists in a procedure that many patients assume is exhaustive.

More important than the raw exposure figures is what they predicted. At the level of individual anatomical sites, the detection rates of early gastric cancer, gastric neoplasms, precancerous conditions, and focal lesions were all significantly higher in adequately exposed sites than in inadequately exposed ones, with differences reaching statistical significance at the level of P less than 0.001. When the analysis moved to the patient level, the pattern held with remarkable consistency: the number of adequately exposed sites per examination correlated positively with lesion detection across every lesion category, with correlation coefficients ranging from 0.830 to 1.000 and all comparisons significant at P less than 0.01. The message is intuitive but now empirically grounded: the more of the stomach lining an endoscopist shows the camera, the more likely the examination is to yield a clinically meaningful finding.

Because simple correlations can be confounded, the investigators went further, applying multivariable logistic regression using generalized estimating equation models that account for clustering of patients within centers. These adjusted models controlled for whether the participant had been allocated to the intervention or control arm of the original trial, the number of standard anatomical sites screened, and the total inspection time. Even after these adjustments, each additional adequately exposed site was associated with higher odds of detecting early gastric cancer, with an adjusted odds ratio of 1.081 and a 95 percent confidence interval of 1.021 to 1.146, statistically significant at P equal to 0.008. The same held for precancerous conditions, with an adjusted odds ratio of 1.056, and for focal lesions, with an adjusted odds ratio of 1.059. For neoplastic lesions overall, the association pointed in the same direction but narrowly missed conventional statistical significance, with an adjusted odds ratio of 1.038 and a P value of 0.060. Sensitivity analyses clustered by individual endoscopist produced materially consistent estimates, strengthening the case that the effect is not an artifact of a few unusually thorough or unusually lax operators.

What makes the OGMEA system potentially transformative is its fit with the existing trajectory of endoscopy quality assurance. Quality indicators have already reshaped colonoscopy, where measures such as adenoma detection rates and cecal intubation rates are now standard benchmarks for training, credentialing, and pay-for-performance programs. Upper gastrointestinal endoscopy has lagged behind, partly because the stomach’s complex, cavernous anatomy makes exposure harder to standardize and partly because no practical tool existed to measure it in real time or in retrospect. The abbreviations that pepper the study, from CCMEA, the cumulative colorectal mucosal exposure area, to GRACE, the gastroscopy rate of cleanliness evaluation, trace the field’s growing effort to borrow quantitative rigor from its lower-gastrointestinal counterpart. An automated measure of gastric exposure, computed by a neural network from the very video stream the endoscope already produces, could be integrated into endoscopy reporting software without adding any burden to the procedure itself, giving departments a continuous, objective feed of quality data.

The study also illustrates a broader shift in how artificial intelligence is being deployed in medicine. Much of the public conversation about AI in endoscopy has centered on computer-aided detection, in which algorithms highlight polyps or suspicious lesions directly. The OGMEA work takes a complementary path: rather than finding the lesion, the algorithm measures whether the human operator created the conditions under which a lesion could be found. This framing sidesteps some of the regulatory and medicolegal complexities of autonomous diagnosis while still harnessing deep learning to improve outcomes. It also generates actionable feedback for trainees, who can see precisely which of the 19 anatomical sites they tend to under-visualize, whether that is the posterior wall of the fundus or the lesser curvature of the antrum, and adjust their technique accordingly.

Cautions remain, as they should with any new quality metric. The exposure thresholds were defined by expert consensus, and different panels might draw the line between adequate and inadequate visualization somewhat differently. The association between exposure and detection, even in adjusted models, is observational in nature; a randomized trial in which endoscopists are compelled to reach OGMEA targets would be needed to prove that raising exposure causally reduces missed cancers. The modest magnitude of the per-site effect, an odds ratio just above 1, accumulates meaningfully only across many sites, which is precisely how endoscopy works in practice but means the metric is best viewed as one component of a broader quality framework rather than a standalone guarantee. Still, with data from more than 25,000 participants across two dozen centers, consistent results at both site and patient levels, and robustness in sensitivity analyses, the OGMEA study makes a compelling case that the stomach’s hidden surfaces deserve to be counted, and that a neural network watching the monitor may soon help ensure that no corner of the gastric mucosa goes unseen.

Subject of Research: Artificial intelligence-based assessment of gastric mucosal exposure during endoscopy and its association with gastric lesion detection

Article Title: Artificial intelligence-based assessment of gastric mucosal exposure and its association with lesion detection: a multicenter study

Article References: Du, H., Chen, Z., Li, J., Zhang, X., Zu, Q., Zeng, Z., Dong, Z., Hassan, C., Yu, H., & Wan, X. (2026). Artificial intelligence-based assessment of gastric mucosal exposure and its association with lesion detection: a multicenter study. BMC Medicine. https://doi.org/10.1186/s12916-026-05288-8

Image Credits: AI Generated

DOI: 10.1186/s12916-026-05288-8

Keywords: artificial intelligence, endoscopy, gastric cancer, mucosal exposure, quality indicators, deep learning, early detection, gastroenterology, precancerous lesions, multicenter study, screening, Artificial

The study, published in Medical & Biological Engineering & Computing, describes a deep learning pipeline built around a hybrid Transformer-convolutional neural network architecture called UDTransNet. The system ingests digitized whole slide images of gastric mucosa biopsies, identifies metaplastic glands at the pixel level, and then translates those segmentation results into a clinical severity grade. Crucially, the grading step follows the Sydney system, the internationally recognized standard that pathologists use to classify gastritis and intestinal metaplasia by estimating the proportion of tissue occupied by metaplastic glands.

The technical challenge the researchers faced is formidable. A single whole slide image can contain tens of thousands of glands scanned at resolutions exceeding a hundred thousand pixels per side, far beyond what any neural network can process in one pass. The team therefore employed multi-scale patch cropping, slicing each slide into overlapping tiles at several magnifications so the network can see both fine cellular texture and broader tissue architecture. UDTransNet then fuses the strengths of two complementary families of deep learning models: Transformers, whose self-attention mechanisms capture long-range spatial relationships across a tissue section, and convolutional neural networks, which excel at extracting precise local edge and texture features. The architecture’s learnable skip connections shuttle these multi-level features back into the decoder, allowing sharp, anatomically faithful segmentation boundaries.

The performance figures reported on the internal test set are striking. The segmentation model achieved a Dice coefficient of 0.9698, meaning that its automatically outlined metaplastic regions overlapped with expert annotations almost perfectly. When those segmentations were converted into Sydney-system severity grades, the model reached an accuracy of 0.8879 with a Kappa statistic of 0.85, a level of agreement conventionally interpreted as near-perfect concordance with expert pathologists.

Perhaps the most provocative comparison in the study is the one against human observers at different career stages. The model’s Kappa of 0.85 exceeded the diagnostic consistency measured between junior and intermediate pathologists, which ranged from 0.67 to 0.82. In practical terms, the algorithm agreed with expert diagnoses more reliably than less experienced human specialists agreed with one another. For clinical workflows in which biopsy volume is high and specialist time is scarce, that gap matters: it suggests a machine assistant could meaningfully reduce the inter-observer variability that has long plagued gastric precancerous lesion grading.

The framework also proved valuable at the other end of the diagnostic spectrum, in ruling disease out. When identifying completely normal tissue, the model achieved a recall of 92.1 percent and an F1 score of 95.9 percent. This ‘negative exclusion’ capability is clinically significant because patients whose biopsies are genuinely free of metaplasia can be triaged toward routine surveillance rather than intensive follow-up, freeing endoscopy and pathology resources for those at genuine risk. An automated system that reliably clears normal slides lets human experts concentrate their attention on the borderline and positive cases where judgment matters most.

What distinguishes this work from many prior AI pathology studies is its interpretability and its direct mapping to a clinical standard. Rather than outputting an opaque risk score, the system produces digital maps of metaplastic gland distribution across the slide and quantifies their area proportion in accordance with Sydney system thresholds. The authors describe this as the first establishment of a digital mapping between pathological morphological features and the Sydney grading criteria. Explanatory visualization techniques such as gradient-based localization, referenced in the paper, allow the model’s attention to be checked against recognizable histological landmarks, addressing a common barrier to clinical adoption of deep learning in medicine.

The clinical context underscores the urgency. Gastric cancer incidence and mortality remain high in East Asia, and the well-characterized Correa cascade of gastric carcinogenesis, from chronic inflammation through atrophy and intestinal metaplasia to invasive cancer, offers a window in which detection and intervention can change outcomes. Global meta-analyses cited by the authors indicate that gastric intestinal metaplasia is widespread, yet its histological grading still depends on subjective visual assessment, a process the study characterizes as inefficient and poorly reproducible. Prior computational efforts have largely focused on endoscopic images; this work moves the diagnostic burden onto the histopathology slide itself, where grading criteria are formally defined.

The implications extend beyond the pathology laboratory. An automated, standardized grading system could harmonize diagnoses across hospitals and countries, enabling more consistent risk stratification for endoscopic surveillance programs and cleaner data for epidemiological research and clinical trials. The authors note that their framework provides an efficient and interpretable solution for intelligent screening of gastric precancerous lesions, and the reported ethics approval from Fujian Provincial Hospital indicates the work is grounded in real clinical specimens. The research was supported by the National Natural Science Foundation of China and provincial science funds of Fujian Province, with contributions from institutions including South China Normal University, Tsinghua University Shenzhen International Graduate School, and Fuzhou University Affiliated Provincial Hospital.

Challenges remain before such systems become routine. External validation on slides from different scanners, staining protocols and patient populations will be essential, and regulatory pathways for AI-assisted diagnosis are still maturing. Yet the core result stands: a Transformer-CNN fusion network, trained on whole slide images and constrained by an internationally accepted grading standard, can match expert pathologists in segmenting and grading one of gastric cancer’s most important precursors. As digital pathology continues to replace glass slides with gigapixel images, tools like this one point toward a future in which every biopsy receives an instant, consistent, and explainable second opinion, one that never tires and never varies from one reading to the next.

Subject of Research: Automated deep learning segmentation and grading of gastric intestinal metaplasia in whole slide pathology images

Article Title: Gastric mucosa intestinal metaplasia segmentation and grading via transformer-CNN fusion architecture: an interpretable digital pathology diagnostic framework

Article References: Jin, Y., Zhu, L., Yan, X., Lin, L., Zhu, M., Yang, L., Han, S., Li, T., Zeng, Y., Ji, Y., & Li, H. (2026). Gastric mucosa intestinal metaplasia segmentation and grading via transformer-CNN fusion architecture: an interpretable digital pathology diagnostic framework. Medical & Biological Engineering & Computing. https://doi.org/10.1007/s11517-026-03658-1

Image Credits: AI Generated

DOI: 10.1007/s11517-026-03658-1

Keywords: gastric cancer, intestinal metaplasia, deep learning, whole slide imaging, digital pathology, semantic segmentation, Sydney system, Transformer, convolutional neural network, precancerous lesions, computer-aided diagnosis, UDTransNet