A vision transformer deep learning model for assessing pediatric ileocolic intussusception severity using ultrasound images

Abstract

Timely identification of children with ileocolic intussusception likely to fail air-enema reduction is critical to avoid delays and bowel perforation. However, even expert sonographers show inter-observer variability. We developed and prospectively validated a Vision Transformer (ViT) deep learning system to predict reduction failure from static B-mode ultrasound images. This multicenter bidirectional cohort study included 5602 children (4–60 months) who underwent air-enema reduction at 14 Chinese tertiary hospitals (retrospective cohort: 2019–2024). After data augmentation, 10,151 images (8122 training, 2029 validation) were used to train a ViT model for binary classification (“success” vs. “failure”). External validation was performed on a prospective cohort of 190 patients (March–June 2025), with three junior and three senior sonographers independently predicting outcomes. The study was approved by the Ethics Committee of Yijishan Hospital of Wannan Medical University (approval No. 2025-04) and registered with ChiCTR2500098673. The model achieved high internal performance (failure: accuracy 0.880, precision 0.969; success: accuracy 0.970, precision 0.898). In the prospective cohort, the ViT model achieved 93.7% overall accuracy, significantly higher than senior (74.7%) and junior (60.7%) sonographers (p < 0.05). This study innovatively applies ViT to assess pediatric ileocolic intussusception severity, providing an objective, accurate tool to support clinical decision-making and reduce treatment risks.

Introduction

Acute intussusception is one of the most common acute abdominal emergencies in children, with ileocolic intussusception being the most frequent type¹. If severe, it can lead to intestinal ischemia, necrosis, perforation, and even death. Timely and accurate diagnosis and treatment are therefore essential². Ultrasound is the preferred diagnostic method, and air-enema reduction under X-ray guidance is the main treatment modality³. However, in cases of severe intussusception or prolonged duration, air-enema may fail, requiring emergency surgery.

Determining the severity of intussusception at the time of presentation to decide whether emergency surgery is needed has long been a clinical challenge for both radiologists and pediatric surgeons^4,5. There is an urgent need for a predictive model to assist in accurate assessment. In recent years, artificial intelligence (AI) has made significant progress in medical imaging, with deep learning algorithms showing excellent performance in image recognition and classification tasks^6–9. However, in the field of pediatric intussusception, AI applications are still in their infancy. Most existing studies focus on basic model development and lack large-scale clinical data analysis. Moreover, most algorithms used are traditional convolutional neural networks (CNNs), which have limitations such as local receptive fields and spatial invariance^10,11.

In 2020, the Vision Transformer (ViT) was introduced as a new deep learning model^12,13. Unlike CNNs, ViT divides images into fixed-size patches and uses self-attention mechanisms to capture global relationships between patches. This gives ViT advantages in global attention, scalability, and cross-modal integration, making it a new paradigm in visual modeling¹⁴.

This study aims to develop a ViT-based deep learning model using ultrasound images to assess the severity of pediatric acute ileocolic intussusception and provide clinical decision support. We collected ultrasound images from 14 hospitals in China between January 2019 and May 2024. The ViT algorithm was used to build and validate the AI model. We then compared the performance of junior, senior, and AI model in a prospective cohort of 190 patients from March to June 2025. The goal is to accurately predict air-enema success and improve clinicians’ ability to assess the severity of intussusception.

Results

Patient characteristics

This study retrospectively collected 5944 ultrasound images from 5602 children diagnosed with acute ileocolic intussusception across 14 centers in China. The number of cases from each hospital is as follows: Yijishan Hospital of Wannan Medical University (962), Hubei Maternal and Child Health Hospital (859), Guangxi Zhuang Autonomous Region Maternal and Child Health Hospital (851), Wuhan Children’s Hospital of Hubei Province (946), Qingdao University Affiliated Hospital (474), Qingdao Women and Children’s Hospital (206), Children’s Hospital Affiliated to Shandong University (74), Liaocheng People’s Hospital of Shandong Province (136), Baoji Maternal and Child Health Hospital of Shaanxi Province (89), Fuyang Women and Children’s Hospital of Anhui Province (247), Anhui Provincial Children’s Hospital (318), Huaibei Maternal and Child Health Hospital of Anhui Province (202), Wanbei Coal and Electricity Group General Hospital Co., Ltd. (138), and Bozhou People’s Hospital of Anhui Province (100). Among the 5602 patients, the median age was 11 months (range: 4–60 months), with 3754 boys (67.01%) and 1848 girls (32.99%). Air-enema reduction was successful in 5163 cases and failed in 439 cases. Additionally, 190 pediatric patients were prospectively enrolled for external validation. Detailed clinical characteristics are presented in Table 1. Figure 1 illustrates the study flowchart.

Table 1.

Datasets for training, validation, prospective pilot study, and Characteristics of patients

Variable	A retrospective dataset (images)			A prospective dataset (patients)
	Total	train_set	valid_set
Final outcome (No. of patients)	5602	4482	1120	190
Enema Success	5163	4131	1032	120
Enema Failure	439	351	88	70
Number of labeled images	5944	4754	1190	190
Enema Success	5287	4229	1058	120
Enema Failure	657	525	132	70
Number of images after data augmentation	10151	8122	2029	190
Enema Success	5552	4442	1110	120
Enema Failure	4599	3680	919	70
Age (Mo.), Median age (range)	11(4–60)	11(4–60)	11(4–60)	10(4–60)
Abdominal Pain time (h), Mean ± SD	10.33 ± 11.60	10.05 ± 9.53	10.94 ± 12.07	10.26 ± 8.78
Temprature (°C), Mean ± SD	36.63 ± 0.58	36.65 ± 0.56	36.58 ± 0.63	36.60 ± 0.69
WBC (×109/L), Mean ± SD	10.25 ± 3.65	10.22 ± 3.79	10.35 ± 3.56	10.19 ± 3.43
Sex, n (%)
Boys	3754(67.01)	3003(67.00)	751 (67.05)	129 (67.89)
Girls	1848 (32.99)	1479(33.00)	369 (32.95)	61 (32.11)
Hematochezia, n (%)
Yes	958 (17.10)	752 (16.78)	206 (18.39)	33 (17.37)
No	4644 (82.90)	3730(83.22)	914 (81.61)	157 (82.63)

Fig. 1. Study flowchart outlining the inclusion, exclusion, and analysis process for both retrospective and prospective cohorts.

A total of 5602 patients were included in the retrospective cohort for model development and internal validation, and 190 patients were enrolled in the prospective cohort for external clinical validation.

Model training, validation, and selection

To address the limited sample size and reduce overfitting, data augmentation was applied, resulting in a total of 10,151 images for model development. The training set comprised 8122 images from 4482 patients, and the validation set included 2029 images from 1120 patients. The Vision Transformer model was trained using these datasets, and the best-performing model was saved based on validation accuracy. Figure 2 shows the training accuracy, training loss, confusion matrix, and class-specific accuracy. The accuracy curves demonstrated steady improvement during training, with both training and validation accuracies approaching 0.95 by the 300th epoch. The model achieved strong classification performance, with high accuracy for both successful and failed reduction cases.

Fig. 2. Training performance of the Vision Transformer model.

a Training accuracy; (b) training loss; (c) confusion matrix on the validation set; (d) class-specific accuracy. The model achieved stable convergence by epoch 300, with high accuracy and low loss on both training and validation datasets.

As shown in Table 2, for predicting air-enema failure, the model achieved an accuracy of 0.8803, precision of 0.9690, recall of 0.8803, and F1-score of 0.9225. For successful reduction, the accuracy was 0.9704, precision 0.8982, recall 0.9740, and F1-score 0.9346. The average accuracy was 0.9291. These results indicate that the model is reliable and clinically applicable. To further evaluate the model’s discriminative ability, calibration, and clinical usefulness, we analysed the independent test set (n = 190). As shown in Table 3 and Fig. 3, the model achieved a sensitivity of 0.886, specificity of 0.967, ROC-AUC of 0.975 (95% CI 0.953–0.991), PR-AUC of 0.962 (95% CI 0.925–0.988), and Brier score of 0.054 (95% CI 0.032–0.079). The calibration curve demonstrated good agreement between predicted probabilities and observed outcomes (Fig. 3C). Decision curve analysis (Fig. 3D) showed that the model provides a positive net benefit over a wide range of threshold probabilities (approximately 0.2–0.9), supporting its potential clinical utility.

Table 2.

Performance of the deep-learning system on the ultrasound images

	Accuracy	Precision	Recall	F1-score
Enema failure	0.8803	0.9690	0.8803	0.9225
Enema success	0.9704	0.8982	0.9740	0.9346
Average acc	0.9291

Table 3.

Performance metrics of the model on the independent test set (n = 190)

Metric	Value (95% CI)
Sensitivity	0.886
Specificity	0.967
ROC-AUC	0.975 (0.953–0.991)
PR-AUC	0.962 (0.925–0.988)
Brier score	0.054 (0.032–0.079)

ROC-AUC area under the receiver operating characteristic curve, PR-AUC area under the precision–recall curve, CI confidence interval. Confidence intervals were estimated using 1000 bootstrap resamples.

Fig. 3. Performance metrics on the independent test set.

Receiver operating characteristic (ROC) and precision–recall (PR) curves of the model for predicting enema failure in the independent test set (n = 190). a ROC curve with area under the curve (AUC) indicating overall discriminative performance. b PR curve with area under the precision–recall curve (PR-AUC) reflecting model performance under class imbalance. c Calibration curve of the model for predicting enema failure in the independent test set (n = 190). The red line represents the apparent calibration, the blue line represents the bias-corrected calibration obtained using bootstrap resampling, and the grey dashed line represents the ideal reference line (perfect calibration). d Decision curve analysis (DCA) for the model predicting enema failure in the independent test set (n = 190). The x-axis represents the threshold probability, and the y-axis represents the net benefit. The model is compared with the treat-all strategy (intervening in all patients) and the treat-none strategy (intervening in no patients). A higher net benefit indicates greater clinical usefulness.

Performance of the deep learning system on ultrasound images

Between March and June 2025, 190 pediatric patients with intussusception were enrolled at four medical centers. Ultrasound images were input into the model, which instantly predicted the probability of air-enema success or failure. Figure 4 illustrates the prospective clinical application of the ViT model. For example, in one case, the model predicted a 97.3% chance of success, and the patient underwent successful air-enema reduction and was discharged uneventfully. In another case, the model predicted only a 3.42% chance of success and a 96.6% chance of failure. The patient indeed failed air-enema reduction and required surgical reduction due to severe intussusception. These examples demonstrate the model’s promising clinical performance and its potential to assist pediatric surgeons and sonographers in assessing intussusception severity and reducing procedural risks.

Fig. 4. Prospective clinical application of the ViT model.

a A case predicted to have a 97.3% probability of success underwent successful air-enema reduction. b A case predicted to have a 96.6% probability of failure indeed failed reduction and required surgical intervention. All images are original to this study. Written informed consent for publication was obtained from the parents of both patients in accordance with institutional ethical approval (No. 2025-04).

Real-world diagnostic performance comparison

To evaluate the model’s generalizability and clinical utility, three junior and three senior sonographers independently interpreted the same set of 190 ultrasound images. Their predictions were compared with those of the AI model.

The prospective test results showed that the average accuracy of junior sonographers was 60.70% (success: 76.94%, failure: 32.86%), while senior sonographers achieved an average accuracy of 74.74% (success: 88.33%, failure: 50.48%). In contrast, the AI model achieved an overall accuracy of 93.68% (success: 96.67%, failure: 88.57%). Supplementary Table 1 shows detailed results. As shown in Table 4, significant differences were observed between each of the 6 sonographers and the AI model (all P < 0.001). The AI model achieved a significantly higher diagnostic accuracy (93.68%) than all individual sonographers, with differences in accuracy ranging from −15.26% to −34.21%. All 95% confidence intervals excluded zero, indicating robust results.

Table 4.

Comparison of interpretation accuracy between each physician and AI model using McNemar test (n = 190)

Comparison Group	Physician Correct Cases	Physician Accuracy (%)	AI Accuracy (%)	Difference in Accuracy (%)	95%CI (%)	McNemar χ²	P-value
Primary sonographers 1 vs. AI	117	61.58	93.68	−32.10	−38.52~ −25.68	51.61	< 0.001
Primary sonographers 2 vs. AI	116	61.05	93.68	−32.63	−38.97~ −26.29	49.69	< 0.001
Primary sonographers 3 vs. AI	113	59.47	93.68	−34.21	−40.43~ −27.99	52.72	< 0.001
Senior sonographers 1 vs. AI	149	78.42	93.68	−15.26	−20.67~ −9.85	18.66	< 0.001
Senior sonographers 2 vs. AI	135	71.05	93.68	−22.63	−28.22~ −17.04	32.44	< 0.001
Senior sonographers 3 vs. AI	140	73.68	93.68	−20.00	−25.26~ −14.74	28.29	< 0.001

Discussion

Acute intussusception in children refers to the telescoping of a segment of intestine and its mesentery into the lumen of an adjacent segment, resulting in intestinal obstruction¹⁵. It is one of the most common acute abdominal emergencies in infants, with ileocolic intussusception being the most frequent type¹⁶. The incidence of intussusception ranks second only to acute appendicitis among pediatric surgical emergencies. If not promptly treated, severe cases can lead to intestinal ischemia, necrosis, perforation, and even death¹⁷. Therefore, timely and accurate diagnosis and treatment are essential to prevent serious complications.

Ultrasound is the preferred diagnostic method due to its non-invasive, real-time, and radiation-free characteristics¹⁸. Air or hydrostatic enema reduction is currently the main non-surgical treatment¹⁹. However, in cases of severe intussusception, enema reduction may fail, and excessive pressure during the procedure can increase the risk of bowel perforation²⁰. Such cases require prompt surgical intervention. In clinical practice, it remains challenging for physicians to accurately assess the severity of intussusception at the time of initial presentation and determine whether emergency surgery is needed.

This study is the first to apply the ViT deep learning model to the assessment of pediatric acute ileocolic intussusception using ultrasound images. Based on a large-scale, multicenter dataset from 14 hospitals in China (5602 patients, 10,151 images), we developed and validated the model using a bidirectional cohort design, combining retrospective modeling with prospective clinical validation. Unlike traditional CNNs, ViT uses a self-attention mechanism to capture global image features, overcoming the limitations of CNNs such as restricted receptive fields and poor modeling of long-range dependencies^21,22. Additionally, we conducted a prospective clinical trial to compare the diagnostic performance of the AI model with that of junior and senior sonographers, further confirming its clinical applicability and generalizability. This study represents not only an algorithmic innovation but also a new technical pathway for precise imaging assessment of pediatric acute abdominal conditions.

Although previous studies have explored the application of AI in pediatric intussusception diagnosis, most have relied on CNN architectures such as YOLOv5, ResNet, or lightweight custom networks, focusing primarily on lesion detection rather than severity assessment^23–26. For example, Kim et al.²⁴ developed a real-time detection system based on YOLOv5 with high accuracy but did not address outcome prediction. Pei et al.²⁵ proposed CIDNet using multicenter data but still employed traditional CNNs, failing to fully utilize global semantic information in images. In contrast, this study introduces ViT into pediatric intussusception ultrasound image analysis for the first time. By dividing images into patches and modeling global dependencies, ViT significantly enhances the model’s ability to express complex lesion features. Its self-attention mechanism can dynamically focus on key regions (e.g., “target sign”, “sleeve sign”) and maintain high discriminative power even in the presence of noise, artifacts, or partial structural blurring. Moreover, this study employed a comprehensive data augmentation strategy combining geometric and photometric transformations, effectively improving model generalizability and addressing issues of sample imbalance and overfitting common in medical image datasets. Taken together, our study surpasses previous reports in terms of model architecture, data processing strategy, and depth of clinical validation.

The ViT model developed in this study achieved an average accuracy of 92.91% in internal validation and 93.68% in prospective testing, significantly outperforming junior (60.70%) and senior (74.74%) sonographers. Especially in identifying high-risk cases of enema failure, the model reached an accuracy of 88.57%, far exceeding human interpretation levels, demonstrating strong clinical auxiliary value. In practice, the model can rapidly assess intussusception severity at admission, assisting clinicians in determining whether air-enema reduction is appropriate, thereby reducing unnecessary procedures and related complications, and improving treatment efficiency. Furthermore, the model outputs probabilistic predictions, offering good interpretability and clinical operability, which facilitates comprehensive decision-making in combination with other clinical indicators. For primary-level hospitals or less experienced clinicians, this AI system can serve as a “virtual senior consultant,” improving diagnostic consistency and accuracy, and holding broad prospects for widespread application. With further optimization and deployment, the model is expected to be integrated into ultrasound devices or PACS systems to enable real-time assisted diagnosis, promoting the intelligence and standardization of pediatric acute abdominal care. In addition, calibration and decision curve analyses demonstrated that the model’s probability estimates are well calibrated (Brier score 0.054) and that the model provides a positive net benefit across clinically relevant threshold probabilities, further supporting its potential utility in guiding clinical decisions. Although formal attention map analysis was not performed in this study, several observations support that our ViT model captures clinically meaningful ultrasound features: consistent performance across 14 hospitals with diverse equipment, robustness to extensive data augmentation, and prospective validation on standardized acquisitions. These findings argue against overfitting to artifacts or site-specific confounders.

Despite these promising results, this study has several limitations. First, only static B-mode ultrasound images were used; color Doppler flow signals and laboratory indicators such as white blood cell count and CRP were not included. The absence of blood perfusion and inflammatory markers may reduce the model’s discriminative power in borderline cases of ischemia or necrosis. We are currently constructing a multimodal dataset that integrates B-mode ultrasound with color Doppler flow parameters (e.g., resistive index of the superior mesenteric artery) and serum biomarkers (WBC, CRP, lactate). A possible fusion strategy is to extract feature embeddings from each modality using separate encoders (ViT for images, MLP for tabular data) and concatenate them before the classification head. Preliminary experiments on a subset of 300 cases show that adding Doppler RI (resistance index > 0.7 as a risk marker) improves failure prediction AUC from 0.965 to 0.981. We plan to prospectively validate a multimodal ViT in the upcoming 500-patient trial. Second, the external validation cohort included only 190 consecutive cases, with 70 positive events (enema failure), resulting in wide confidence intervals. A larger multicenter non-inferiority trial (target ≥ 500 cases, ≥ 150 failure events) has been registered and is expected to be completed by 2026 to validate the model’s stability and safety in a broader population. Third, the current PyTorch implementation averages 470 ms per image on an RTX 3080 GPU, which does not yet meet the < 200 ms requirement for real-time bedside decision-making. The main bottlenecks are the Transformer self-attention computation (~60% of inference time) and image preprocessing/post-processing. We plan to optimize the model using TensorRT/ONNX for INT8 quantization and graph optimization, and test deployment on edge devices such as the NVIDIA Jetson Orin Nano, targeting ≤150 ms latency and ≤ 1 GB memory usage for integration into portable ultrasound devices. Preliminary optimization results have already reached ~180 ms on the Jetson Orin Nano, and further improvements are ongoing. Fourth, the failure rate differed substantially between the retrospective (7.8%) and prospective (36.8%) cohorts. This is primarily explained by case-mix shift: the prospective cohort was enrolled exclusively from four regional high-level pediatric referral centers that routinely manage a much higher proportion of complex, severe, and previously failed cases, whereas the retrospective cohort reflected the typical tertiary hospital population (failure rate 5–10%). Despite this, the ViT model maintained excellent discriminative performance, demonstrating robustness to case-mix variation. However, users should be aware that the expected failure prevalence and positive predictive value will change with the target population, and may require threshold calibration when deploying the model in lower-risk settings.

We also acknowledge the potential risks of false positives (FP) and false negatives (FN). FP cases—where the model over-predicts enema failure—may lead to unnecessary general anesthesia and surgery, increasing psychological and financial burden on families. FN cases—where high-risk patients are missed—may delay surgery and increase the risk of bowel necrosis, perforation, and shock. Therefore, we propose a “dual-signature” workflow:

1. AI provides a probability score;

2. A junior physician reviews the image—if AI and physician agree, the corresponding plan is executed;

3. If there is disagreement or the AI confidence is between 0.4 and 0.6, a senior physician is automatically consulted for final decision-making.

Based on decision curve analysis and calibration curves on the validation set, we observed that when the model’s predicted probability fell in the 0.4–0.6 interval, the accuracy dropped to ~75%, and approximately 30% of all misclassifications occurred within this range. By threshold sweeping, we found that setting a low-risk threshold at 0.4 (i.e., probability < 0.4 suggests direct trial of air-enema reduction) and a high-risk threshold at 0.6 (probability > 0.6 suggests preparation for surgery or senior physician review) could reduce the risk of misclassification to < 5%, while keeping the proportion of cases requiring manual review low (~15% of all cases). A simulation of the dual-signature workflow on the retrospective data showed that if cases with AI confidence in the 0.4–0.6 range were referred to senior physicians for decision-making, the overall accuracy would increase to 96.2%, with no false-negative case (failure predicted as success) leading to severe adverse outcomes. We plan to formally validate the safety of this workflow in a future prospective trial.

This process leverages AI’s screening efficiency while retaining expert oversight, significantly reducing misdiagnosis rates and aligning with the rapid management requirements of pediatric acute abdominal conditions. Future prospective trials will compare the “dual-signature” and “AI-only” models to quantify the real-world benefits of human–AI collaboration.

Conclusion

In summary, this study innovatively applies the Vision Transformer deep learning model to the assessment of pediatric acute ileocolic intussusception severity. The proposed model provides an objective and accurate tool to assist clinical decision-making and reduce treatment risks. Compared with traditional diagnostic methods, the ultrasound-based AI model can more accurately predict the success rate of air-enema reduction, offering valuable support to clinicians and reducing misdiagnosis and missed diagnosis due to inexperience or subjective factors. Furthermore, the use of multicenter data greatly enhances the model’s generalizability and reliability, making it suitable for pediatric patients across different regions. Nevertheless, the prospective validation cohort in this study was limited to 190 patients; larger and more diverse cohorts are needed to confirm these preliminary findings before widespread clinical adoption.

Methods

Data collection

This study retrospectively collected 5944 ultrasound images from 5602 children diagnosed with acute ileocolic intussusception across 14 centers in China. These hospitals included Yijishan Hospital of Wannan Medical University, Hubei Maternal and Child Health Hospital, Guangxi Zhuang Autonomous Region Maternal and Child Health Hospital, Wuhan Children’s Hospital of Hubei Province, Qingdao University Affiliated Hospital, Qingdao Women and Children’s Hospital, Children’s Hospital Affiliated to Shandong University, Liaocheng People’s Hospital of Shandong Province, Baoji Maternal and Child Health Hospital of Shaanxi Province, Anhui Provincial Children’s Hospital, Huaibei Maternal and Child Health Hospital of Anhui Province, Fuyang Maternal and Child Health Hospital of Anhui Province, Wanbei Coal and Electricity Group General Hospital Co., Ltd., and Bozhou People’s Hospital of Anhui Province.

This study retrospectively collected 5944 ultrasound images from 5602 children diagnosed with acute ileocolic intussusception across 14 centers in China. These hospitals included Yijishan Hospital of Wannan Medical University, Hubei Maternal and Child Health Hospital, Guangxi Zhuang Autonomous Region Maternal and Child Health Hospital, Wuhan Children’s Hospital of Hubei Province, Qingdao University Affiliated Hospital, Qingdao Women and Children’s Hospital, Children’s Hospital Affiliated to Shandong University, Liaocheng People’s Hospital of Shandong Province, Baoji Maternal and Child Health Hospital of Shaanxi Province, Anhui Provincial Children’s Hospital, Huaibei Maternal and Child Health Hospital of Anhui Province, Fuyang Maternal and Child Health Hospital of Anhui Province, Wanbei Coal and Electricity Group General Hospital Co., Ltd., and Bozhou People’s Hospital of Anhui Province. A total of 190 pediatric patients were prospectively enrolled from four of these centers (Yijishan Hospital of Wannan Medical University, Hubei Maternal and Child Health Hospital, Guangxi Maternal and Child Health Hospital, and Wuhan Children’s Hospital) between March and June 2025 for external validation (independent test set).

Inclusion criteria were: children diagnosed with acute ileocolic intussusception by ultrasound; aged 4–60 months; and received air-enema reduction under X-ray guidance as the initial treatment. Exclusion criteria were: small bowel or colonic intussusception; initial treatment with hydrostatic reduction under ultrasound guidance; and presence of other severe comorbidities. All diagnoses were confirmed by senior pediatric surgeons and radiologists based on clinical history and ultrasound findings. After admission, all patients underwent air-enema reduction performed by experienced pediatric surgeons. Data collected included patient demographics, medical history, ultrasound images, and treatment outcomes. Data integrity and accuracy were ensured to support high-quality model development.

This study was approved by the Ethics Committee of Yijishan Hospital of Wannan Medical University (approval No. 2025-04) and registered with the Chinese Clinical Trial Registry (ChiCTR2500098673). Informed consent was obtained from the parents or legal guardians of all patients included in the prospective study. For retrospective data, patients were informed at admission that their clinical data might be used for research, and all data were de-identified before analysis.

Model development

Ultrasound images were collected at initial presentation before any intervention. Only images with typical features such as the “target sign” or “pseudokidney sign” were included. All images were reviewed and confirmed by a team of three senior pediatric surgeons and two pediatric radiologists. All splits were performed at the patient level before any data augmentation. Specifically, patients were randomly assigned to training and validation sets in an approximately 4:1 ratio, and all images (including augmented versions) from a given patient were confined to the same set, ensuring no leakage between training and validation.

Preprocessing steps included image resizing to 224 × 224 pixels using bicubic interpolation, noise reduction with median and Gaussian filtering, and image enhancement using contrast-limited adaptive histogram equalization (CLAHE). Data augmentation was applied to increase dataset size and improve generalizability, using geometric transformations (rotation ± 15°, translation ± 10 pixels, scaling ±10%, horizontal/vertical flipping) and photometric adjustments (brightness, contrast, saturation, hue).

A professional team of pediatric radiologists and sonographers annotated the images. Lesion boundaries were delineated, and images were labeled as “success” or “failure” based on final air-enema outcomes. A second senior doctor reviewed all annotations to ensure consistency and accuracy. We calculated inter-annotator agreement among three senior annotators (two pediatric radiologists and one pediatric surgeon) on a randomly selected subset of 500 images. For outcome label assignment, the agreement rate was 98.4% (Fleiss’ kappa = 0.96). For lesion bounding boxes, the mean intersection-over-union (IoU) was 0.82. Figure 5 presents the overall pipeline of the AI system, including data augmentation, model training, and validation.

Fig. 5. Overall pipeline of the AI system, including data augmentation, model training, and validation.

The system processes ultrasound images through preprocessing, patch embedding, Transformer encoding, and classification to predict air-enema reduction outcomes.

The ViT model divides the input image into fixed-size non-overlapping patches. Each patch is flattened and linearly embedded into a lower-dimensional vector. These embedded vectors, along with positional encodings, are fed into a Transformer encoder that consists of alternating multi-head self-attention (MHSA) and feed-forward layers, with layer normalization and residual connections. After the Transformer encoder, the output is passed through a multilayer perceptron (MLP) classification head to predict the probability of air-enema success or failure (Fig. 6).

Fig. 6. Architecture of the Vision Transformer (ViT)–based deep learning model for classifying pediatric ileocolic intussusception outcomes using static B-mode ultrasound images.

Input images are divided into fixed-size patches, linearly embedded, and processed through a Transformer encoder with multi-head self-attention. The final MLP head outputs probabilities for air-enema success or failure.

Hardware and software specifications are listed in Table 5. Data were input in batches (size 16 or 32). During each forward pass, the model calculated the loss between predicted and true labels, and backpropagation was used to update model parameters. Training continued for up to 300 epochs or until validation performance plateaued. Hyperparameters such as learning rate, batch size, and regularization were optimized based on validation results. To address class imbalance (5163 success vs. 439 failures in the retrospective cohort), we employed weighted cross-entropy loss with a class weight of 5 for the failure class and applied class-balanced sampling to oversample failure cases in each mini-batch. Weighted cross-entropy was found to outperform Focal Loss in our preliminary experiments.

Table 5.

Experimental setup and specifications

Component	Specification
CPU	Intel Core i7-6800K 3.5Ghz
RAM	16GB DDR4 3600 MHz
GPU	Nvidia RTX3080 11GB
Operating System	Ubuntu 16.04
Deep Learning Framework	Pytorch

Model evaluation was performed using accuracy, precision, recall, F1-score, AUC, confusion matrix, and Cohen’s Kappa coefficient. The best-performing model was selected based on validation accuracy and F1-score. In addition, we calculated sensitivity, specificity, the area under the receiver operating characteristic curve (ROC-AUC), and the area under the precision-recall curve (PR-AUC) for predicting enema failure on the independent test set (n = 190). Calibration was assessed using a calibration plot (observed vs. predicted probabilities) and the Brier score (lower values indicate better calibration). Decision curve analysis (DCA) was performed to evaluate the clinical net benefit of the model across a range of threshold probabilities, comparing the model with “treat-all” and “treat-none” strategies. All 95% confidence intervals for performance metrics were estimated using 1000 bootstrap resamples.

This study did not train or compare any alternative deep learning baseline model (e.g., ResNet, YOLO, or other CNN architectures). The primary objective was to evaluate the feasibility and clinical utility of the ViT for classifying air-enema outcomes from pediatric intussusception ultrasound images, and to prospectively compare its performance with that of junior and senior sonographers. Thus, no inter-model comparison (e.g., ViT vs. ResNet) was performed.

Clinical validation

After model development, clinical validation was conducted to assess real-world performance. An independent internal test set was used to evaluate model performance on unseen data. Predictions were compared with final clinical outcomes to calculate performance metrics and analyze misclassification patterns.

For prospective testing, 190 consecutive patients were enrolled from March to June 2025. Three junior sonographers (<10 years of experience) and three senior sonographers (> 10 years) independently interpreted ultrasound images and predicted air-enema outcomes. AI model predictions were generated in real time, and all predictions were compared with final clinical results.

A standardized test protocol was developed with participating departments. Junior sonographers received training to ensure consistency. Ultrasound images were transmitted to the AI model in real time, and clinical data (symptoms, history, lab results) were also recorded. Performance of AI and human readers was compared using the same metrics.

To facilitate interpretation of model performance across phases, we systematically documented differences in equipment, operators, and patient selection between the retrospective and prospective cohorts. The retrospective cohort (14 hospitals) used various ultrasound devices (e.g., GE Logiq E9, Philips EPIQ 7, Siemens Acuson S2000) operated by certified sonographers from each hospital. Acquisition protocols followed the Chinese guidelines for pediatric intussusception ultrasound examination but showed minor inter-center variations. In contrast, the prospective cohort (4 centers) uniformly employed two device types (GE Logiq E9 and Philips EPIQ 7). All prospective scans were performed by sonographers with ≥5 years of pediatric ultrasound experience, following a standardized acquisition protocol that specified fixed scan planes, depth, and gain settings. Patient selection was consistent across both phases: the retrospective cohort included all consecutive eligible children diagnosed between 2019 and 2024, while the prospective cohort comprised consecutive eligible patients enrolled from March to June 2025, applying the identical inclusion and exclusion criteria.

Statistical analysis

Data were analyzed using SPSS 26.0 and R 4.2.0. Normality was assessed using the Shapiro–Wilk test. Categorical variables were compared using the Chi-square or Fisher’s exact test. Continuous variables were compared using Student’s t-test or Mann–Whitney U test, as appropriate. To compare the classification performance of the artificial intelligence model with that of individual ultrasound physicians in the same group of 190 prospective cases, we used McNemar’s paired binary result test (correct/incorrect prediction). A two-tailed p-value < 0.05 was considered statistically significant.

Author contributions

All authors have approved the final version of the manuscript. J.L., Y.W. and D.Z. contributed equally to this paper. J.L., Y.W., D.Z., M.L. and M.L. designed the study with support from all the co-authors. M.L. led the development of the method with. M.L. and Y.W. wrote the code and carried out the experiments. X.M., Y.Y., F.H., W.X., Q.D., D.Y., Z.J., P.Z., Z.Z., S.Q., Z.L., H.Q., Y.C., H.W., P.W., Z.L., L.B., K.S., Z.D., A.L., G.S., T.Z., G.W., Q.H., L.H., W.M. and H.L. provided ultrasound images and diagnostic labeling. J.L., Y.W, H.L. and M.L. analyzed and interpreted the results. J.L. and Y.W. provided technical and material support. J.L., Y.W., H.L. and M.L. were involved in drafting the paper. All authors contributed to the article and approved the submitted version.

Data availability

The ultrasound image data used in this study are not publicly available due to patient privacy protection policies of the participating hospitals. De identified data may be made available to qualified researchers upon reasonable request to the corresponding author, subject to institutional review board approval. The custom code developed for this study is not publicly available. It may be made available to qualified researchers upon reasonable request to the corresponding author.

Code availability

The custom code developed for this study is not publicly available. It may be made available to qualified researchers upon reasonable request to the corresponding author.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41746-026-02941-7.