The clinical problem the researchers set out to solve is deceptively simple to state but difficult to solve in practice. Dynamic transperineal ultrasound, or TPUS, is a noninvasive imaging method in which a probe is placed on the perineum, the area between the vagina and the anus, to visualize the bladder, urethra, and surrounding pelvic structures. During the examination, the patient is asked to perform the Valsalva maneuver, forcefully exhaling against a closed airway as if straining, which increases abdominal pressure and causes pelvic organs to descend. The sonographer must then identify the single most informative frame from the resulting video, trace the position of the urethrovesical junction, the point where the urethra meets the bladder, and grade the degree of descent according to the Green classification system, a scheme first proposed in 1975 that divides cystocele severity into types I, II, and III based on the angle and orientation of the urethra relative to the bladder base.

Every one of those steps depends on human judgment. Selecting the peak Valsalva frame requires recognizing the moment of maximal descent amid a noisy, rapidly changing image sequence. Placing landmarks on the urethrovesical junction demands anatomical expertise, and small errors in landmark placement can shift a patient from one Green type to another. Studies of pelvic floor ultrasound have long documented substantial variability between observers, and the classification is further complicated by confounders such as levator co-activation, in which involuntary contraction of the pelvic floor muscles during straining masks the true extent of organ descent. The result is that Green classification, despite its clinical value in guiding surgical planning for anterior vaginal wall prolapse, remains operator-dependent in a way that many other ultrasound measurements do not.

The Green-AttGRU system was designed to compress that entire manual workflow into an automated pipeline. The architecture combines a deep convolutional neural network for visual feature extraction with a gated recurrent unit, a type of recurrent neural network well suited to sequential data, augmented with an attention mechanism that allows the model to focus on the most diagnostically relevant frames in the ultrasound video. In practical terms, the network watches the whole Valsalva sequence the way a sonographer would, learns which frames capture the moment of maximal bladder neck descent, localizes the urethrovesical junction, and outputs a Green type classification without any manual frame selection or landmark tracing. The name reflects this design: the attention-enhanced gated recurrent unit sits at the heart of the classification engine.

To train and validate the system, the team assembled a dataset of 881 patients examined at Shengjing Hospital, a tertiary referral center. Of these, 688 patients formed a retrospective development cohort used to train the model, while 193 patients were enrolled prospectively and formed an independent test cohort that the model had never seen during training. This separation matters enormously in machine learning for medicine, because models that are evaluated only on the data they were trained on routinely overstate their performance. The prospective design, in which patients were enrolled and scanned after the model architecture was fixed, provides a more honest estimate of how the system would behave in clinical use. The study was approved by the hospital’s ethics committee and conducted in accordance with the Declaration of Helsinki, with written informed consent obtained from prospective participants.

The standalone performance of Green-AttGRU on the complete prospective test cohort was strong. The model achieved a macro-averaged area under the receiver operating characteristic curve, a measure of diagnostic discrimination across all severity classes, of 0.939, with a 95 percent confidence interval of 0.897 to 0.971. Its overall accuracy was 0.902, meaning it agreed with the reference standard in roughly nine out of ten cases. For a three-class classification task performed on dynamic ultrasound video, where the boundaries between Green types are defined by continuous anatomical angles that patients can sit near, those numbers place the automated system in the range of what experienced human readers can achieve, but with the crucial advantage of perfect repeatability: the same input always produces the same output.

The more clinically revealing experiment was the reader study, which asked whether the AI could make human radiologists better rather than simply replace them. Four radiologists, two junior and two intermediate in experience, independently interpreted 67 prospective patient examinations under two conditions: unaided, and with the AI system’s output available as decision support. Without AI assistance, the four readers achieved overall accuracies ranging from 0.761 to 0.821, with macro-averaged F1 scores, which balance precision and recall across classes, between 0.660 and 0.777. With the AI’s classification available, accuracy rose to a range of 0.851 to 0.881, and macro-F1 climbed to 0.820 to 0.860. The improvement was consistent across readers, suggesting that the benefit was not confined to the least experienced members of the panel.

Perhaps the most striking findings concerned agreement and speed. Before AI assistance, the four radiologists agreed with one another only moderately, with a Fleiss’ kappa, a statistic that measures inter-rater agreement beyond chance, of 0.453. After consulting the AI, that figure jumped to 0.786, indicating substantial agreement. In other words, the system did not just make the readers more accurate; it made them more consistent with one another, converging on a shared interpretation of ambiguous scans. At the same time, the pooled median interpretation time per case fell from 26.7 seconds to 9.9 seconds, a reduction of more than 60 percent. For a busy pelvic floor imaging service, that difference compounds quickly, and it points to a workflow benefit that goes beyond diagnostic quality alone.

The study’s authors are careful to frame these results as evidence of preliminary feasibility rather than proof of readiness for unsupervised clinical deployment. The data come from a single tertiary referral hospital, and the reader study involved eight radiologists in total across two experience levels, a sample that cannot capture the full spectrum of expertise and scanning conditions found in the wider clinical world. The system was also designed and evaluated for one specific task, Green classification of the anterior compartment, whereas a complete pelvic floor ultrasound assessment involves additional measurements, including hiatal dimensions and organ descent at rest and on straining, that the current pipeline does not address. External validation at multiple centers, with different scanner hardware and patient populations, remains the necessary next step before any regulatory or guideline body would consider routine use.

Even with those caveats, the work adds to a rapidly growing body of evidence that deep learning can standardize the interpretation of pelvic floor ultrasound, a field that has historically lagged behind obstetric imaging in automation. Recent studies have demonstrated deep learning models for identifying pelvic floor organs in the midsagittal plane, for automating the evaluation of female pelvic organ descent, and for enhancing three-dimensional transperineal ultrasound biometry in prolapse assessment. What distinguishes the present study is its end-to-end scope, spanning frame selection, landmark localization, and classification within a single workflow-oriented system, and its prospective reader study design, which follows the kind of evaluation framework that clinical prediction model reporting guidelines now recommend. If subsequent multicenter studies replicate these results, AI-assisted Green classification could become a practical tool for reducing variability in pelvic floor imaging, shortening examination times, and ultimately helping clinicians choose the right surgical approach for the millions of women whose bladder support fails them.

Subject of Research: Development and prospective validation of an AI-assisted clinical decision support system for automated Green classification of cystocele on dynamic transperineal ultrasound.

Article Title: An AI-assisted Clinical Decision Support System for Green Classification of Cystocele on Dynamic Transperineal Ultrasound

Article References: Zhu, H., Geng, X., Zhou, H., Guo, W., Dai, Y., Zhang, H., Dong, M., Li, H., & Wang, X. (2026). An AI-assisted Clinical Decision Support System for Green Classification of Cystocele on Dynamic Transperineal Ultrasound. Journal of Medical Systems, 50(1), Article 127. https://doi.org/10.1007/s10916-026-02453-7

Image Credits: AI Generated

DOI: 10.1007/s10916-026-02453-7

Keywords: cystocele, transperineal ultrasound, Green classification, artificial intelligence, clinical decision support, pelvic organ prolapse, deep learning, reader study, urethrovesical junction, Valsalva maneuver, radiology, pelvic floor

The study stands out for a methodological choice that most clinical AI deployments still lack. Conventional binary classifiers — the workhorses of medical machine learning — produce a single probability for every case, and a fixed threshold converts that probability into a decision of malignant or benign. Such systems never signal doubt. They output a number even when the input image is ambiguous, the acquisition quality is poor, or the case falls far outside anything they were trained on. The research team, led by first authors Hang Ling, Cailing Lin and Jing Ning, with corresponding author Ziwei Zhang, built their framework around a different premise: that a diagnostic algorithm should be able to abstain. Their selective triage rule combines three ingredients — a fusion prediction model, a composite uncertainty score, and a pre-locked routing policy that sends confident cases down automatic pathways and routes uncertain ones to senior human review.

The technical architecture is worth unpacking. The fusion model integrates three distinct information streams: structured clinical data, features extracted from multimodal ultrasound imaging, and variables mined from structured ultrasound reports. Multimodal ultrasound is an important component here, since modern neck ultrasound is not a single image but a constellation of modalities — grayscale morphological features such as echogenicity and border characteristics, Doppler vascular patterns, and in some settings contrast-enhanced ultrasound dynamics. By fusing these with clinical context and report-derived information, the model aims to approximate the holistic judgment a seasoned sonographer applies, rather than relying on pixel patterns alone.

Just as important is what the researchers did with uncertainty. Rather than trusting the model’s raw confidence, they constructed a composite uncertainty measure from two complementary signals: predictive entropy, normalized against its distribution in the development cohort, and probability dispersion estimated through bootstrap resampling of the model’s fits. Entropy captures how peaked or flat the model’s output distribution is on a given case, while bootstrap dispersion captures how sensitive the prediction is to perturbations in the training data — a proxy for how far the case sits from the model’s comfort zone. Combining the two, the team fixed an uncertainty threshold of U = 0.700. Critically, every free parameter was locked during development: the binary classification threshold at p = 0.537, the selective-routing probability boundaries at p = 0.320 and p = 0.660, and the uncertainty cutoff. Once locked, the entire pipeline was applied to the internal-validation and two external-validation cohorts without any retuning whatsoever — a design that guards against the subtle overfitting that plagues many retrospective AI studies.

The study population comprised 518 patients, with one index lymph node analyzed per person: 206 in the development cohort, 88 in internal validation, and 112 in each of two independent external cohorts. Discrimination remained remarkably stable across sites, a finding that in itself deserves attention because performance degradation at external sites is the most common failure mode of published clinical prediction models. The area under the receiver operating characteristic curve was 0.839 in internal validation, 0.850 in external validation cohort 1, and 0.842 in external validation cohort 2. At the locked binary threshold, sensitivity and specificity were 0.595 and 0.902 in internal validation, 0.657 and 0.911 in the first external cohort, and 0.817 and 0.808 in the second. Those numbers describe a competent but not extraordinary classifier — which is precisely the point, because the selective framework was engineered to compensate for the model’s fallibility rather than to hide it.

The heart of the paper lies in its selective-triage results. In the pooled external-validation population of 224 patients, 38 were routed to a lower-risk automatic pathway, 44 to a higher-risk automatic pathway, and 142 — more than sixty percent — were deferred for senior review. Automatic coverage was therefore 0.366, while selective accuracy among the automated cases reached 0.927, with a 95 percent confidence interval of 0.849 to 0.966. The accepted errors within the automated subset were small in absolute terms: two false negatives and four false positives. Among patients funneled into the lower-risk automatic pathway, the negative predictive value was 0.947, meaning the residual probability of malignancy in that group was 5.3 percent — a figure the authors report transparently with a wide confidence interval spanning roughly 1.5 to 17.3 percent. The higher-risk pathway achieved a positive predictive value of 0.909. A risk–coverage analysis, summarized by a partial area under the risk–coverage curve of 0.037 in the pooled external data, quantified how selective accuracy behaved as coverage was expanded or contracted.

The authors are unusually explicit about the limits of these numbers. Only about 37 percent of external cases were eligible for automatic routing, and the accepted false-negative and false-positive counts, while modest, are not zero. Two of the 38 patients automatically assigned to the lower-risk pathway turned out to have malignant nodes. In a separate per-case analysis, the composite uncertainty score was only moderately effective at flagging binary-model errors, achieving an area under the curve of 0.581 for error detection and an area under the precision-recall curve of 0.247 — numbers that indicate real room for improvement in the uncertainty estimation itself. The team concludes that the framework should be interpreted as a retrospective, research-stage selective-routing demonstration rather than an established clinical safety or workflow tool, a disclaimer that rare in a field where press releases routinely outrun the evidence.

Where the system showed its most immediately practical benefit was in the eight-reader study. Using crossed reader-by-case bootstrap resampling — the gold-standard multi-reader multi-case methodology for interpreting diagnostic accuracy studies — the researchers tested whether access to the model’s output changed reader performance. Junior readers improved their accuracy by 0.055, a statistically significant gain with a confidence interval of 0.022 to 0.089 and a p-value of 0.002. Middle-level and senior readers showed no statistically significant change, a pattern consistent with the intuition that the tool functions as a form of expert guidance for less experienced practitioners while adding little for those who already possess the pattern-recognition skills it encodes. If the framework ultimately translates to the clinic, its clearest value proposition may be compressing the training gap between junior and senior diagnosticians, rather than replacing expert judgment outright.

The broader significance of the study lies in its modeling of what responsible clinical AI could look like. Rather than chasing headline accuracy figures, the researchers built their entire evaluation around the questions that actually matter for deployment: when should an algorithm be allowed to act autonomously, how much residual risk is acceptable within the automated zone, and how much workload is genuinely deferred to humans. The reported numbers answer those questions soberly. Roughly a third of cases could be automated with a combined accuracy above 92 percent, but the framework’s own honesty mechanisms pushed nearly two-thirds of patients toward senior review, and even the automated pathway retained a nonzero malignancy risk. The team also adhered to modern reporting and risk-of-bias standards, referencing the TRIPOD+AI and PROBAST+AI frameworks, and the study received ethics approval from Fujian Provincial Hospital with center-specific authorization for the external cohorts. Funded by the Natural Science Foundation of Fujian Province, the work offers a template — conservative, externally validated, and uncertainty-aware — for a generation of diagnostic AI systems whose most important capability may be knowing what they do not know.

Subject of Research: An uncertainty-aware selective artificial intelligence triage model for distinguishing benign from malignant cervical lymphadenopathy on multimodal ultrasound

Article Title: An uncertainty-aware ultrasound triage model for cervical lymphadenopathy: a retrospective multicenter development and external validation study

Article References: An uncertainty-aware ultrasound triage model for cervical lymphadenopathy: a retrospective multicenter development and external validation study. (n.d.). https://doi.org/10.1186/s12880-026-02780-8

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02780-8

Keywords: cervical lymphadenopathy, ultrasound, artificial intelligence, uncertainty quantification, selective prediction, multimodal imaging, external validation, triage, diagnostic AI, reader study, medical imaging, clinical decision support