Diagnostic Performance of Multimodal Large Language Models for Grading Pediatric Vesicoureteral Reflux on MCUG: A Preliminary Study

ABSTRACT

Aims:

To evaluate the diagnostic accuracy of ChatGPT-5 and Gemini 2.5 Pro compared with a pediatric radiologist reference standard in grading vesicoureteral reflux (VUR) on pediatric micturating cystourethrogram (MCUG) studies.

Materials and Methods:

This retrospective study included 125 pediatric MCUG cases, comprising 25 cases per VUR Grade (I-V) as determined by radiologists. A balanced distribution of VUR grades was employed to facilitate a comprehensive assessment of model performance across the severity spectrum. For each case, the most representative image was anonymized, and the stored Joint Photographic Experts Group images were randomly uploaded to ChatGPT-5 and Gemini 2.5 Pro with a standardized prompt requesting VUR grading. Grades assigned by the models were compared with the radiologist’s consensus standard.

Results:

VUR was present in 182/250 (72.8%) renal units, more frequently on the left side (80%) than the right side (65.6%). Both models demonstrated the highest accuracy in identifying the absence of reflux (ChatGPT-5 = 92.6% and Gemini 2.5 Pro = 88.2%). Overall diagnostic accuracy was 52% (131/250) for ChatGPT-5 and 46% (115/250) for Gemini 2.5 Pro. ChatGPT-5 performed best for grades III–IV (51% and 49%, respectively), whereas Gemini 2.5 Pro performed best for Grade V (54%). Agreement with radiologists was fair (ChatGPT-5 κ = 0.40; Gemini 2.5 Pro κ = 0.33).

Conclusion:

ChatGPT-5 and Gemini 2.5 Pro demonstrated fair but suboptimal accuracy in grading VUR. Model performance appears to be influenced by the prompting strategy used. Multimodal large language models show potential as adjunctive tools; however, further optimization of prompting approaches and multimodal integration is required before clinical application.

INTRODUCTION

Artificial intelligence (AI) is rapidly transforming radiology by enhancing data processing, report generation, and resident education.[1,2] However, its integration into pediatric radiology remains limited.[3,4] Among AI applications, multimodal large language models (MLLMs) such as ChatGPT (OpenAI, San Francisco, CA) and Gemini (Google, Mountain View, CA) have shown promise in assisting clinical reporting and decision support.[5,6,7,8,9,10] Large language models are AI systems trained primarily on text data to generate and interpret human-like language. In contrast, MLLMs extend these capabilities by integrating multiple data modalities, including images, enabling simultaneous processing of visual and textual inputs. With recent advances in image-based training, these models are being increasingly applied to the interpretation of radiologic images. Yet, most studies have focused on adult datasets, leaving their role in pediatric imaging largely unexplored.[11] There is a notable gap in the literature regarding the application of MLLMs to pediatric fluoroscopic imaging, which requires dynamic, temporally dependent interpretation rather than static image analysis. An important factor contributing to the scarcity of studies in pediatric radiology is the limited availability of openly accessible pediatric imaging datasets.

Evidence regarding the use of MLLMs in pediatric radiology, particularly for routine fluoroscopic procedures such as Micturating cysturethrogram (MCUG), is scarce. MCUG remains the standard study for evaluating the urinary tract and diagnosing vesicoureteral reflux (VUR) – the retrograde flow of urine from the bladder into the ureters or upper urinary tracts. High-grade VUR predisposes to recurrent pyelonephritis and renal scarring, necessitating accurate grading for optimal management.[12,13]

VUR grading is subject to known interobserver variability despite the use of standardized grading systems. This limitation has prompted growing interest in AI-based approaches aimed at improving objectivity and reproducibility in VUR assessment.

Evaluating VUR grading using MLLMs will provide insight into their potential for radiologic image interpretation. Grading VUR on MCUG is particularly challenging for AI because interpretation relies on both anatomical recognition and temporal changes during reflux. Because these proprietary models are easily accessible to nonexperts, it is essential to acknowledge the limitations and potential inaccuracies that may arise when such tools are used without professional oversight. Identifying these knowledge gaps can guide future strategies – either by refining model training through expert-supervised datasets or by increasing public awareness regarding the limitations of AI-based image interpretation.

Therefore, this study aimed to evaluate the diagnostic accuracy of ChatGPT-5 and Gemini 2.5 Pro compared with a pediatric radiologist reference standard in grading VUR on pediatric MCUG studies.

MATERIALS AND METHODS

Study design

This retrospective diagnostic accuracy study evaluated the ability of two commercially available AI-based MLLMs – ChatGPT-5 (OpenAI) and Gemini 2.5 Pro (Google) – to grade VUR on pediatric MCUG studies, using interpretation by a pediatric radiologist as the reference standard. The study was performed in September 2025, and paid subscriptions were taken for ChatGPT-5 and Gemini 2.5 Pro. The study was approved by the Institutional Review Board, and the requirement for informed consent was waived because all data were anonymized and contained no identifiable patient information.

Study sample

A total of 125 pediatric MCUG studies were included, comprising 25 cases of each of the five VUR grades (I–V). A balanced distribution of VUR grades was employed to facilitate a comprehensive assessment of model performance across the severity spectrum, acknowledging that this design may not reflect real-world prevalence. Cases were randomly retrieved from the departmental picture archiving and communication system. Cases with complete MCUG studies and adequate visualization of the VUR were included, whereas those with incomplete demographic information or unclear clinical indication were excluded. For each study, demographic details, clinical indication, and the reference VUR grade were recorded.

Reference standard

VUR grading for each MCUG study was performed by one of three pediatric radiologists, each with more than 10 years of experience in pediatric radiology. Each case was independently evaluated by a single radiologist using a standardized reference grading scheme; no consensus reading was performed, and interobserver variability was not assessed.

Grading was performed according to the American College of Radiology–Society for Pediatric Radiology practice parameter for voiding cystourethrography in children.[14] The reference grading scheme is summarized in Table 1. Grades III–V were considered higher grades.

Table 1.

Grading of vesicoureteral reflux

Grade	Description
I	Reflux into ureter only
II	Reflux into ureter, pelvis, and calyces without dilatation
III	Mild or moderate dilatation/tortuosity of ureter with mild or moderate dilatation of pelvis and calyces; fornices mildly blunted
IV	Moderate dilatation/tortuosity of ureter and pelvis with blunting of fornices but preserved papillary impressions
V	Gross dilatation and tortuosity of ureter and pelvis with loss of papillary impressions in most calyces

Multimodal large language model-based image interpretation

A single representative MCUG image per case was selected for MLLM analysis. Representative images for each case were selected by a pediatric radiology fellow using a standardized criterion to ensure consistency across cases. The selected image corresponded to the phase of peak reflux, demonstrating maximal ureteric and pelvicalyceal dilatation, representing the highest grade of reflux in the index case. Image selection was not blinded to the final diagnosis. As MLLMs are sensitive to visual artifacts, we used standardized images obtained directly from fluoroscopic acquisition, without any cropping or annotations. Image resolution and contrast were uniform across the dataset. After anonymization, images were stored in Joint Photographic Experts Group format [Figure 1].

Figure 1.

Representative fluoroscopic images from micturating cystourethrogram (MCUG) demonstrating grades of vesicoureteral reflux according to the American College of Radiology–Society for Pediatric Radiology classification: (a) Grade I – reflux into ureter only; (b) Grade II – reflux into ureter, pelvis, and calyces without dilatation; (c) Grade III – mild to moderate ureteral and pelvicalyceal dilatation with minimal blunting; (d) Grade IV – moderate dilatation and tortuosity with blunting of fornices; (e) Grade V – gross dilatation and tortuosity with loss of papillary impressions

Prompt design

These images were randomly uploaded to both ChatGPT-5 and Gemini 2.5 Pro with the standardized prompt. The prompt used was as follows:

“Identify the grade of VUR on each side.”

The prompt did not include any definition of the VUR grading. Only a single prompt was tested, and the robustness of that prompt was not evaluated.

Model output assessment

The response generated by the models also included the grading used. The grading used by MLLMs matched the grading used by the pediatric radiologist. In addition to assigning a grade to the image, MLLMs also generated an explanation of the grading. Bilateral VUR was analyzed at the renal unit level, with each side modeled independently. Figures 2 and 3 illustrate the responses generated by ChatGPT-5 and Gemini 2.5 Pro to the prompt.

Figure 2.

Example output generated by ChatGPT-5 for a pediatric MCUG image demonstrating bilateral Grade V vesicoureteral reflux (VUR). The model was provided a standardized prompt (“Identify the grade of VUR on each side”) and produced both a grade classification and explanatory reasoning

Figure 3.

Example output generated by Gemini 2.5 Pro for a pediatric MCUG image demonstrating bilateral Grade V vesicoureteral reflux using the same standardized prompt. The response includes assigned grades and accompanying interpretative explanation

Grades assigned by both models were recorded and compared with the radiologist reference standard to determine diagnostic accuracy and interobserver agreement. Qualitative assessment was also performed to identify common interpretive errors made by the models.

Figure 4 summarizes the entire method in a flowchart form.

Figure 4.

Flowchart depicting the study design

Statistical analysis

Statistical analyses were performed using IBM SPSS Statistics, version 31.0 (IBM Corp., Armonk, NY, USA). Continuous variables were summarized as mean ± standard deviation, and categorical variables as frequencies and percentages. Diagnostic accuracy, sensitivity, specificity, negative predictive value (NPV), and positive predictive value (PPV) for each MLLM were calculated relative to the radiologist consensus grades. Agreement between MLLMs and radiologists was evaluated using Cohen’s κ (coefficient), interpreted as: slight (<0.20), fair (0.21–0.40), moderate (0.41–0.60), substantial (0.61–0.80), and almost perfect (0.81-1.00). PPV and NPV are prevalence-dependent measures and, given the artificially balanced case distribution in this study, may not be generalizable to real-world clinical settings.

RESULTS

Demographic profile and clinical indication

A total of 125 pediatric patients (250 renal units) were included for analysis. The cohort comprised 86 boys (68.8%) and 39 girls (31.2%), with a mean age of 5.3 ± 3.7 years (range, 0.75–18 years; median, 5 years).

The most common indication for performing MCUG was follow-up or surveillance imaging in previously diagnosed cases on conservative management or after intervention (n = 41, 33%). Other indications included suspected or treated posterior urethral valves (PUVs) (n = 25, 20%), recurrent urinary tract infections (n = 16, 13%), and hydronephrosis (n = 13, 10%).

Vesicoureteral reflux grade distribution

Each renal unit was assigned a VUR grade (I–V) by a pediatric radiologist. VUR was detected in 182 renal units, with the following overall grade distribution:

• Grade I: 32 kidneys

• Grade II: 35 kidneys

• Grade III: 41 kidneys

• Grade IV: 39 kidneys

• Grade V: 35 kidneys

• No reflux: 68 kidneys.

This distribution is a consequence of the study design, which included a predefined number of cases across VUR grades, and does not represent the true prevalence of VUR in the clinical population. Bilateral reflux was observed in 57 patients (45.6%). VUR occurred more frequently on the left side (80%) compared with the right side (65.6%). The detailed sidewise grade distribution is summarized in Table 2.

Table 2.

Distribution of grade of reflux on right and left side

VUR grade	Right side, n (%)	Left side, n (%)
No reflux	43 (34.4)	25 (20)
I	14 (11.2)	18 (14.4)
II	11 (8.8)	24 (19.2)
III	22 (17.6)	19 (15.2)
IV	15 (12)	24 (19.2)
V	20 (16)	15 (12)

VUR: Vesicoureteral reflux

Overall, higher-grade reflux (grades III–V) was more common on the left side (53.6%) than on the right side (45.6%).

Performance of ChatGPT-5 and Gemini Pro 2.5

Anonymized representative images of each case were uploaded to both models. ChatGPT-5 correctly identified 131 of 250 renal units (52.4%), while Gemini 2.5 Pro correctly identified 115 of 250 renal units (46%).

• Right side: ChatGPT-5, 72/125 (57.6%); Gemini 2.5 Pro, 62/125 (49.6%)

• Left side: ChatGPT-5, 59/125 (47.2%); Gemini 2.5 Pro, 53/125 (42.4%).

Among 68 renal units without reflux, ChatGPT-5 correctly classified 63 (92.6%), and Gemini 2.5 Pro correctly classified 60 (88.2%).

Diagnostic accuracy

The overall diagnostic accuracy of ChatGPT-5 was 52% (95% confidence interval [CI]: 46.1%–58.7%), while Gemini 2.5 Pro had an overall diagnostic accuracy of 46% (95% CI: 39.8%–52.2%). It is summarized in Table 3.

Table 3.

Overall diagnostic accuracy of ChatGPT-5 and Gemini 2.5 Pro

	ChatGPT-5 (95% CI)	Gemini 2.5 pro (95% CI)
Sensitivity (%)	37.3 (30.2–44.9)	30.2 (23.8–37.4)
Specificity (%)	92.6 (83.3–97.5)	88.2 (77.0–94.9)
PPV* (%)	93.1 (88.3–96.1)	87.3 (80.9–92.0)
NPV* (%)	35.6 (28.4–43.5)	32.1 (25.2–39.8)
Overall diagnostic accuracy	52 (46.1–58.7)	46 (39.8–52.2)

*PPV and NPV are prevalence-dependent measures and, given the artificially balanced case distribution in this study, may not be generalizable to real-world clinical settings. PPV: Positive predictive value, NPV: Negative predictive value, CI: Confidence interval

Both models showed lower sensitivity (37.3% for ChatGPT vs. 30.2% for Gemini 2.5 Pro), but higher specificity (92.6% for ChatGPT-5 vs. 88.2% for Gemini 2.5 Pro). The PPV was higher (ChatGPT-5 = 93.1%, Gemini 2.5 Pro = 87.3%) than the NPV (ChatGPT-5 = 35.6%, Gemini 2.5 Pro = 32.1%).

When adjusted for a disease prevalence of 1%,[15] both models demonstrated a marked decline in PPV with preservation of high NPV. For ChatGPT-5, the PPV decreased to 4.9%, and the NPV increased to 99.3%. Similarly, for Gemini 2.5 Pro, the PPV decreased to 2.5%, and the NPV increased to 99.2%. These findings indicate that at low prevalence, both models function better as rule-out tools rather than rule-in tests.

Figure 5 represents the diagnostic performance of both AI models.

Figure 5.

Bar chart comparing diagnostic performance of ChatGPT-5 and Gemini 2.5 Pro in grading vesicoureteral reflux on MCUG at the renal unit level. Metrics include sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and overall diagnostic accuracy. Error bars represent 95% confidence intervals. PPV and NPV are dependent on the study sample distribution and may not reflect real-world prevalence

Grade-wise performance and common misclassifications

No reflux: ChatGPT-5 correctly identified the absence of reflux in 63 renal units out of 68 renal units. It falsely identified reflux in 5 renal units and mislabeled them incorrectly as grades I, II, III, and IV. Two cases were mislabeled as grade III, while 1 each was graded as I, II, and IV.

Gemini 2.5 Pro correctly identified the absence of reflux in 88% cases (n = 60), while eight cases were false positives. It mislabeled 3 cases each as Grade IV and V, while 1 case each was mislabeled as Grade I and III.

Grade I (n = 32): ChatGPT-5 correctly identified only 19% (n = 6) of these cases while mislabeling 50% of these cases as no reflux (n = 16). 6 cases (19%) were upgraded to Grade II.

Gemini 2.5 Pro had 28% (n = 9) accuracy for Grade I. It missed Grade I reflux completely in as high as 66% of cases (n = 21), mislabeling them as no reflux. It labeled 2 of the cases as Grade IV and did not predict any Grade II or III.

Grade II (n = 35): There was 20% (n = 7) accuracy for Grade II by ChatGPT-5. 43% (n = 15) of these cases were assigned to the no reflux category. It overestimated some of these cases in higher grades; for example, 4 cases were mislabeled as Grade V.

Gemini 2.5 Pro had only 14% (n = 5) accuracy in its prediction. It overestimated 12 cases (34%) as Grade IV and mislabeled 5 cases (14%) as Grade V. In addition, it misidentified 10 cases (29%) as having no reflux.

Grade III (n = 41): ChatGPT-5 had the highest accuracy in identifying Grade III cases, as 51% (n = 21) were correctly identified. In 13 (32%) of these cases, it was upgraded to Grade IV. Five of these cases were completely missed, as no reflux, while one case was graded as VUR Grade I.

The Gemini 2.5 Pro had low accuracy in identifying Grade III. It correctly identified only 15% of cases (n = 6). It mislabeled 49% of cases (n = 20) as Grade IV. Five cases were mislabeled as having no reflux, while eight cases were mislabeled as Grade II.

Grade IV (n = 39): ChatGPT-5 correctly identified 49% (n = 19) of these cases. It often mislabeled 10 cases (26%) as Grade III, while over-grading 6 cases (15%) to Grade V. Four cases were misidentified as no reflux.

Gemini 2.5 Pro correctly identified 41% (n = 16) of these cases. It mislabeled 11 cases (28%) as Grade V, while 6 cases (15%) were undergraded as Grade II. Reflux was completely missed in 3 cases, which were mislabeled as no reflux.

Grade V (n = 35): ChatGPT-5 correctly identified 43% (n = 15) of these cases. In the remaining cases, eight cases (23%) were mislabeled as Grade IV, two cases (6%) as Grade III, and seven cases (20%) as Grade II. Two cases were assigned as no reflux.

The Gemini 2.5 Pro had the highest accuracy in identifying these grades. It correctly identified 54% (n = 19) of these cases. 37% (n = 13) of cases were downgraded as Grade IV, and 3 (9%) cases were mislabeled as Grade III. It did not mislabel Grade V as Grade I and II, and no reflux.

Common misclassification with grade-wise accuracy is summarized in Table 4.

Table 4.

Common misclassification and grade-wise diagnostic accuracy of ChatGPT-5 and Gemini 2.5 pro

True grade	ChatGPT-5	Gemini 2.5 pro	Diagnostic accuracy of ChatGPT-5 (%)	Diagnostic accuracy of Gemini 2.5 Pro (%)
No reflux	Grades I–IV (mostly grade III)	Grade I, III–V (mostly IV–V)	93	88
Grade I	No reflux (50%), Grade II (19%)	No reflux (66%), Grade IV (6%)	19	28
Grade II	No reflux (43%), Grade V (11%)	No reflux (29%), Grade IV–V (48%)	20	14
Grade III	Grade IV (32%)	Grade IV (49%), Grade II (20%)	51	15
Grade IV	Grade III (26%), Grade V (15%)	Grade V (28%), Grade II (15%)	49	41
Grade V	Grade IV (23%), Grade II–III (26%)	Grade IV (37%), Grade III (9%)	43	54

Confusion matrix analysis demonstrated that ChatGPT frequently misclassified low-grade reflux (grades I–II) as no reflux, whereas Gemini 2.5 Pro showed a tendency to overestimate severity, particularly misclassifying grades II–III as Grade IV [Figure 6].

Figure 6.

Confusion matrices comparing model-predicted vesicoureteral reflux grades with radiologist reference standards at the renal unit level. (a) ChatGPT-5 demonstrates frequent underestimation of low-grade reflux, particularly misclassifying grades I–II as no reflux. (b) Gemini 2.5 Pro shows a tendency toward overestimation, with frequent misclassification of intermediate grades (II–III) as Grade IV. Rows represent true grades and columns represent predicted grades. Numbers within each cell indicate counts of renal units

Degree of agreement

There was a fair degree of agreement between MLLMs and the radiologist reference standard, with Cohen’s κ = 0.40 for ChatGPT-5 and κ = 0.33 for Gemini 2.5 Pro.

DISCUSSION

We evaluated the diagnostic accuracy of two AI-based MLLMs – ChatGPT-5 and Gemini 2.5 Pro – for grading VUR in pediatric patients. Both models demonstrated lower diagnostic accuracy than a pediatric radiologist, with only fair inter-observer agreement. Misclassification patterns were similar across grades for both models. Our cohort primarily comprised young children (median age, 5 years) at high risk for VUR, as reflected by the clinical indications for MCUG. A majority (63%) of reflux grades were high-grade (III–V), representing clinically significant disease, while mild cases were also included.

ChatGPT-5 achieved an overall diagnostic accuracy of 52%, characterized by high specificity but limited sensitivity, indicating reliable detection of absent reflux yet difficulty in recognizing low-grade disease. Nearly half of the Grade I–II cases were misclassified as normal. Grades III and IV were frequently interchanged, and approximately one-fourth of Grade V cases were underestimated.

Gemini 2.5 Pro performed slightly less accurately (46%) with comparable specificity (88%). It showed the greatest difficulty distinguishing grades II–III and a tendency to overestimate reflux severity. Despite lower accuracy, it was less likely than ChatGPT-5 to label true reflux as absent.

Accurate VUR grading is essential for management decisions.[16,17] However, given the moderate accuracy and fair agreement observed, these models are not currently suitable for clinical use. Their potential role as adjunctive tools remains dependent on significant improvements in accuracy and further validation. VUR is inherently a dynamic process, and its grading on micturating cystourethrography (MCUG) is optimally performed through real-time assessment of fluoroscopic sequences. Reflux may be intermittent and may not reach its maximum grade simultaneously in both ureters, particularly in low-grade cases. In the present study, the models were evaluated using single static images rather than dynamic fluoroscopic sequences. This approach does not capture the temporal evolution of reflux or side-specific variability, which are integral to accurate grading in clinical practice. Consequently, this may have contributed to grading inconsistencies and could have influenced the overall diagnostic performance of the models.

VUR grading is known to have interobserver variability. In our study, both models struggled most with low-grade reflux (grades I–II), likely reflecting the inherent difficulty in detecting subtle ureteral dilation and minimal pelvicalyceal involvement on static images. Unlike higher grades, where gross ureteral tortuosity and calyceal blunting provide more conspicuous visual cues, low-grade reflux is characterized by transient and often subtle findings that may be difficult to appreciate without dynamic fluoroscopic assessment. The use of a single representative image in our study further limits the ability to capture the temporal evolution of reflux, which is critical for accurate grading in borderline cases. ChatGPT-5 more frequently underestimated disease, often misclassifying low-grade reflux as no reflux, whereas Gemini 2.5 Pro demonstrated a tendency toward overestimation, possibly reflecting differing internal thresholds for abnormality. Both models struggled to distinguish between Grade III and IV, consistent with known human interobserver variability. Prior studies by Craig et al.[18] and O’Neil et al.[19] reported excellent interobserver agreement (κ >0.9), whereas Schaeffer et al.[20] found moderate interobserver agreement (κ = 0.66–0.68) for grading of VUR among radiologists. In comparison, our κ values of 0.40 (ChatGPT-5) and 0.33 (Gemini 2.5 Pro) indicate substantial underperformance relative to human observers. This also highlights the scope for improvement in future models to better detect Grade III versus IV.

The existing literature supports the utility of MLLMs for text-based radiology tasks, but evidence for their use in image-based interpretation, particularly in children, remains limited.[21,22] Reith et al. conducted a study evaluating the performance of MLLMs (ChatGPT-4, Gemini 1.5 Pro, and Claude 3 Opus) in interpreting pediatric radiology images.[23] They assessed performance for 90 cases across different anatomical systems. It was found that all models correctly identified only 25 cases with an overall diagnostic accuracy of 27.8%. ChatGPT-4 and Gemini 1.5 Pro had an overall equal diagnostic accuracy of 23.3%. They used two cases of PUV and one case of VUR. ChatGPT-4 was unable to identify each of these cases, whereas Gemini 1.5 Pro could partially diagnose PUV and correctly diagnose VUR. However, direct comparison with our findings is limited by important differences in study design, including the use of heterogeneous case types, a small number of genitourinary cases (only one VUR case), and evaluation of earlier model versions.

Similarly, Atakır et al. evaluated the diagnostic performance of ChatGPT-4o version in radiological practice.[24] They evaluated a total of 129 cases in which a combination of radiological images only, radiological images with clinical data, and radiological images with clinical data and diagnostic options was presented to ChatGPT-4o. ChatGPT-4o performed poorly when only image options were provided, with an accuracy of 19.9%, whereas its accuracy was higher (80.8%) when presented with clinical information, diagnostic options, and radiological images. These results are not directly comparable to our study, which used a uniform image-only approach with a single standardized prompt and focused specifically on pediatric MCUG interpretation.

Taken together, these differences in patient population, task specificity, model versions, and prompting strategy limit direct comparability but collectively suggest that model performance is highly context-dependent.

Our study indicates that the latest versions of ChatGPT-5 and Gemini 2.5 Pro have better accuracy than their predecessors. However, there is still scope for improvement in these models, as the diagnostic accuracy does not match that of a trained pediatric radiologist.

We would like to highlight that these models are general-purpose MLLMs and not medical imaging devices. Unlike FDA-cleared software trained on large, domain-specific visual datasets, they are not trained for VUR, are not optimized for pediatric fluoroscopic images, and operate without DICOM metadata. Therefore, their performance in this study should not be interpreted as diagnostic inadequacy, but rather as evidence of their misapplication to tasks beyond their intended scope.

The clinical utility of MLLMs in grading VUR on MCUG extends beyond diagnostic automation and has direct implications for pediatric surgical decision-making. Accurate differentiation between low- and high-grade reflux is pivotal, as management strategies diverge substantially between conservative surveillance and surgical intervention, particularly in borderline Grade III–IV cases. Variability in MCUG interpretation is well recognized, especially in assessing ureteral dilatation, calyceal blunting, and pelvic tortuosity. MLLMs demonstrating high agreement with expert pediatric radiologists could reduce interobserver variability, standardize grading, and provide a reproducible second-read mechanism. From a surgical perspective, such consistency may improve patient selection for endoscopic injection therapy or ureteric reimplantation, reduce unnecessary interventions in low-grade reflux, and facilitate timely escalation in high-grade disease associated with recurrent febrile urinary tract infections or renal scarring. Furthermore, the multimodal capability to integrate imaging findings with clinical parameters – such as infection history and prior functional imaging – positions these systems as potential decision-support tools rather than mere image classifiers. In this context, MLLMs may be better positioned as adjunctive decision-support tools, assisting radiologists by providing a preliminary interpretation, highlighting potential findings, and promoting consistency, rather than functioning as independent diagnostic systems. If externally validated across institutions, MLLMs could enhance multidisciplinary workflow efficiency, support evidence-based surgical planning, and improve long-term renal preservation in affected children.

Our study had a few limitations. These MLLMs are general-purpose models trained on a corpus of texts rather than medical images. These models are continually improving in their diagnostic abilities; however, this study can serve as a reference for future research. We acknowledge that a convenient sample size was used in this study; however, the results may vary when applied to a natural clinical population. Future studies can be planned to utilize these models, in which they are tested prospectively, and the VUR grade is reported before the final assignment of grading by an expert pediatric radiologist. This will better gauge the real-world applicability. The objective of our study was to evaluate the performance of general-purpose MLLMs rather than specialized DL classifiers; we relied on a single standardized prompt and a single inference per case. Prompt variation, structured grading instructions, and exploration of hyperparameter settings can meaningfully influence model performance in MLLMs. Future work should incorporate prompt refinement strategies such as structured prompts with explicit grading criteria, integration of clinical context, and multi-step or iterative prompting to evaluate improvements in model performance.

We acknowledge that, in the present study, preselected representative images depicting peak reflux were used. This removes an important decision-making step typically performed by the radiologist and may simplify the diagnostic task for the models. Consequently, the reported model performance may not accurately reflect real-world clinical settings, where dynamic assessment and frame selection are required.

CONCLUSION

Under the specific prompting strategy used in this study, both ChatGPT-5 and Gemini 2.5 Pro demonstrated fair but suboptimal accuracy in grading VUR, a key determinant of clinical management. However, these findings should be interpreted with caution, as model performance is highly dependent on prompt design and optimization. As prompt robustness and alternative prompting strategies were not systematically evaluated, it would be premature to conclude that these models are inherently unsuitable for VUR grading. Future studies incorporating structured prompt engineering and sensitivity analyses are essential to better define the true diagnostic potential of these systems.

Conflicts of interest

There are no conflicts of interest.

Funding Statement

Nil.