The diagnostic value of radiomics based on two-dimensional ultrasound in staging diabetic nephropathy

Abstract

Objective

To stage diabetic nephropathy (DN) using two-dimensional ultrasound (B-mode) radiomics combined with clinical features.

Methods

DN was classified into early, middle, and late stages. Two-dimensional ultrasound images and clinical biochemical data from patient records were analyzed. Radiomics features were extracted from images, and two classification scenarios were examined: early vs. middle to late DN, and early to middle vs. late DN. Lasso logistic regression was used to create nomograms integrating clinical and radiomics data. The performance of these nomograms was evaluated using ROC curves, calibration, and decision curves.

Results

242 patients with renal biopsy (early DN: n = 102; middle DN: n = 53; late DN: n = 87) were included and randomly split into training (n = 169) and validation (n = 73) sets. For early vs. middle to late DN, the nomograms achieved AUCs of 0.939 and 0.876, with sensitivities of 0.882 and 0.816, specificities of 0.896 and 0.686, and F1 scores of 0.905 and 0.775 in training and validation cohorts, respectively. For early to middle vs. late DN, AUCs were 0.951 and 0.955, with sensitivities of 0.767 and 0.889, specificities of 0.917 and 0.913, and F1 scores of 0.800 and 0.873, respectively. Decision curve analysis confirmed the superiority of the combined model.

Conclusion

Nomograms based on ultrasound radiomics and clinical features effectively distinguish DN stages non-invasively.

Clinical trial number

Not applicable.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12880-026-02422-z.

Introduction

Diabetic nephropathy is a severe microvascular complication affecting approximately 40% of diabetic patients, with 31.3% of them progressing to end-stage renal disease as the disease advances [1, 2]. Against the backdrop of a global diabetes prevalence of 564 million, projected to rise to 600 million by 2035, it poses a significant threat to patient health [3].

The staging of diabetic nephropathy (DN) is closely associated with effective treatment strategies and long-term prognosis. Research indicates that early diagnosis of DN can reduce the risk of progression to end-stage renal disease by 80% [4]. Currently, glomerular filtration rate and proteinuria are the main criteria for staging DN. However, in clinical practice, it is commonly observed that some DN patients may have more severe renal damage even in mild abnormalities in GFR or microalbuminuria, requiring kidney biopsy for staging confirmation [5]. Therefore, finding a non-invasive and accurate method to reflect DN staging has become a direction for improving the prognosis of DN patients.

Microalbuminuria and estimated glomerular filtration rate (eGFR) are key items recommended for screening in the clinical practice guidelines for managing type 2 diabetes mellitus (T2DM) [6]. However, studies have shown that eGFR and microalbuminuria in T2DM patients may not reflect disease progression accurately [7, 8]. Assessing the severity and prognosis of diabetic nephropathy (DN) based solely on these two parameters is evidently insufficient. Two-dimensional ultrasound measurements provide information such as kidney size and shape, renal capsule, internal echoes, and cortical and medullary thickness. Research indicates that ultrasound has some discriminatory ability among different phenotypes of DN [9]. However, two-dimensional ultrasound is not sensitive enough to detect early changes in kidney size or echo alterations. In summary, these indicators alone cannot efficiently and accurately diagnose the staging of DN, thus the urgent need to find effective combined methods to achieve higher value and potential in the diagnosis of DN staging.

Radiomics refers to the use of machine learning methods to extract and analyze features from medical images, exploring anatomical and functional information of disease pathology. These high-throughput data can reduce the workload of clinical professionals and improve the quality of diagnosis, prognosis, and disease management [10, 11]. Currently, radiomics has been widely applied in distinguishing between benign and malignant renal masses [12], determining the size, staging, and grading of renal cell carcinoma [13], analyzing stone composition [14], and detecting fibrosis in research [15]. Our preliminary research found that radiomics based on two-dimensional ultrasound has certain value in diagnosing early diabetic nephropathy [16]. However, the role of radiomics based on two-dimensional ultrasound in the staging of diabetic nephropathy is still unclear.

Based on the above reasons, we try to build an applied model and verify the model by combining the two-dimensional kidney ultrasound and clinical biochemical data most commonly performed in the daily physical examination of patients with diabetic nephropathy, so as to provide a non-invasive diagnostic method for disease staging in patients with diabetic nephropathy.

Materials and methods

The study followed the Declaration of Helsinki and was approved by the hospital ethics committee (2023252). As this study was retrospective, we waived the signing of written informed consent.

Selection of study participants

Patients with diabetic nephropathy and an eGFR < 60 mL/min/1.73 m² were included in the study. Inclusion criteria: renal biopsy, specific kidney stage data; Exclusion criteria: missing ultrasound data or biochemical data; low ultrasound image resolution, low development quality; patients with renal tumors, renal stones, and renal cysts. Ultrasound images and biochemical data were collected in the same time window, and the time difference between them did not exceed 2 days.

Pathological staging of diabetic nephropathy

All patients underwent renal biopsy prior to inclusion. Biopsy specimens were fixed in formalin, embedded in paraffin, and stained with hematoxylin-eosin, periodic acid-Schiff (PAS), Masson’s trichrome, and periodic acid-silver methenamine (PASM). Pathological staging of diabetic nephropathy was performed according to the Renal Pathology Society (RPS) classification [17]. Based on this system, patients were classified into three groups: Early DN: Class I (glomerular basement membrane thickening) and Class IIa (mild mesangial expansion); Middle DN: Class IIb (severe mesangial expansion) and Class III (nodular sclerosis – Kimmelstiel-Wilson lesions); Late DN: Class IV (advanced diabetic glomerulosclerosis).

Ultrasound image acquisition

All two-dimensional ultrasound images were acquired prior to renal biopsy using commercially available ultrasound systems (Logiq P6, GE Healthcare, USA). A convex array transducer with a frequency range of 2–8 MHz was used for all examinations. The imaging protocol was standardized as follows: patients were examined in the supine or lateral decubitus position; the depth was adjusted to visualize the entire kidney; gain settings were optimized to ensure clear visualization of renal contours and internal architecture; harmonic imaging was enabled to improve image quality. All examinations were performed by two experienced radiologists (with 5 and 8 years of experience in renal ultrasound, respectively). Images were stored in DICOM format and exported to the radiomics platform for further analysis. Due to the retrospective nature of this study, minor variations in individual acquisition parameters across patients may exist, which represents a limitation of this work.

Processing flow of radiomics

Image preprocessing: All ultrasound images were resampled to a uniform voxel spacing to ensure consistency in feature extraction. Image normalization was applied to minimize the impact of acquisition-related intensity variations.

Feature standardization: All extracted radiomics features were normalized using Z-score transformation prior to feature selection and model construction, ensuring that features with different scales contribute equally to the analysis.

Image segmentation

We imported the ultrasound images in DICOM format into Darts and Discovery Platform for image annotation((https://carevault.neusoft.com/darts), we annotated the whole kidney as a region of interest, and any disagreement in it was resolved by group discussion.

Feature extraction and establishment of radiomics label

The Darts and Discovery Platform can not only annotate regions of interest but also extract radiomics features using its integrated AI toolkit. (https://carevault.neusoft.com/discovery). The retrieved features are both first-order (histogram and morphological features) and second-order parameters. The second-order parameters primarily consist of the grayscale co-occurrence matrix (GLCM), grayscale run length matrix (GLRLM), grayscale size region matrix (GLZSM), adjacent gray adjustment difference matrix (NGTDM), and grayscale correlation matrix. Two sonographers described the ROI for all photos, and alternate ICC values above 0.75 were used for further analysis. A total of 1,025 radiomics features were extracted from each ultrasound image, including first-order statistics, shape-based features, and texture features (GLCM, GLRLM, GLSZM, NGTDM, and GLDM). To assess feature reproducibility, two sonographers independently delineated the ROIs on all images, and inter-observer intra-class correlation coefficients (ICCs) were calculated for each feature. Features with ICC < 0.75 were considered unstable and excluded from further analysis. After ICC-based filtering, 64 features (6.2%) with ICC < 0.75 were excluded, and the remaining 961 features (93.8%) with ICC ≥ 0.75 were retained for subsequent analysis. Patients will be randomly assigned to the training and validation groups in a 7:3 ratio. We intend to conduct two types of comparisons: Early DN VS middle to late stages DN and Early to middle DN VS late stage DN. The rationale for choosing these two binary classification tasks was based on clinical relevance. In clinical practice, physicians face two key binary decisions at different stages of DN progression: First, distinguishing early-stage disease from progressive disease (mid-to-late DN) to guide intensification of therapy; And second, distinguishing non-advanced (early-to-middle) from advanced (late) disease to prepare for renal replacement therapy. These binary questions align directly with real-world treatment decisions. Methodologically, binary classification offers greater robustness with modest sample sizes and provides more intuitive probability estimates for clinically meaningful outcomes compared to a three-class model.

Feature selection

We employed LASSO regression to construct a predictive model for diabetic nephropathy staging. LASSO proves effective by automatically selecting crucial features, shrinking less significant ones towards zero via a tuning parameter λ determined via 10-fold cross-validation. This approach enhances model simplicity and predictive accuracy by prioritizing the most impactful variables.

Model construction and result validation

We used the Neusoft Discovery platform for machine learning analysis to construct a logistic regression model as follows (https://carevault.neusoft.com/discovery): Firstly, the dataset was split into a training cross-validation set (trainCV-set) and a test set in an 70 − 30% ratio. Secondly, within the trainCV-set, we performed ten iterations of ten-fold cross-validation (CV). Each iteration involved splitting the training set into nine parts for training and one part for validation. We evaluated the logistic regression model’s mean square error (MSE) on the CV set and selected the model with the MSE closest to the average as the final model. Thirdly, the performance of the final logistic regression model was evaluated on an independent test set, including metrics such as AUC, sensitivity, specificity, and F1 score. These steps ensure the model’s robustness during training, validate it rigorously through cross-validation, and assess its effectiveness in predicting diabetic nephropathy staging through independent testing. Additionally, we conducted evaluations using calibration curves and decision curve analysis (DCA). DCA determines the model’s clinical utility by calculating net benefit at various probability thresholds.

Statistical analyses

Statistical testing was performed using SPSS (version 26.0; IBM) and R statistical software (version 4.0.2). A p-value < 0.05 was considered statistically significant. The normality of the data was determined using the Kolmogorov-Smirnov test. The Student’s t-test was used to compare normally distributed continuous variables, with results given as mean and standard deviation. Non-normally distributed variables were evaluated with the Wilcoxon rank-sum test and reported as quartiles. The chi-square test was used to compare categorical variables, which were reported as frequencies.

Results

Basic clinical Information

A total of 242 patients with biopsy-confirmed diabetic nephropathy (DN) were included in this study, comprising 102 early-stage, 53 middle-stage, and 87 late-stage DN patients. The patient selection flowchart is presented in Fig. 1, and baseline clinical characteristics stratified by DN stage are summarized in Table 1. Patients were randomly divided into training (n = 169, 70%) and validation (n = 73, 30%) cohorts. Table 2 shows no significant differences in clinical characteristics between the two cohorts, indicating successful randomization. The overall radiomics workflow is illustrated in Fig. 2.

Fig. 1.

Flow diagram of the included participants

Table 1.

Patient’s characteristics at baseline

Item	Early DN (n = 102)	Middle DN (n = 53)	Late DN (n = 87)		p
Age, y	57(50–62)	59(50–67)	59(53–65)		0.204
Gender, n (man/female)	73/29	35/18	60/27		0.773
CHO	4.69(3.72–5.81)	4.97(4.32–6.64)	4.83(4.00-5.64)		0.166
TG	1.82(1.31–2.97)	1.83(1.13–2.52)	1.78(0.95–2.51)		0.178
HDL-C	0.96(0.77–1.15)	1.03(0.88–1.27)	1.02(0.78–1.31)		0.188
LDL-C	2.85(1.93–3.80)	3.11(2.60–4.23)	2.78(1.97–3.49)		0.048
Apo-A	1.16(1.01–1.35)	1.23(1.05–1.40)	1.14(0.93–1.37)		0.104
Apo-B	1.01(0.76–1.35)	1.09(0.88–1.52)	1.01(0.82–1.46)		0.119
APB: APA	0.85(0.66–1.10)	0.93(0.74–1.18)	0.91(0.66–1.22)		0.438
BUN	7.01(5.22–9.86)	9.49(6.89–14.24)	17.89(11.66–23.60)		< 0.001
CREA	105.50(73.00-161.32)	166.20(99.30–348.00)	533.00(269.00-745.00)		< 0.001
BUN/CREA	0.06(0.05–0.08)	0.06(0.04–0.07)	0.04(0.03–0.05)		< 0.001
UA	405.89 ± 109.82	416.66 ± 126.52	432.77 ± 112.75		0.278
GLU	7.46(5.54–9.86)	7.29(5.71–10.32)	6.53(4.96–7.80)	0.022

Abbreviations: DN, Diabetic Nephropathy; CHO, Total Cholesterol; TG, Triglycerides; HDL-C, High-Density Lipoprotein Cholesterol; LDL-C, Low-Density Lipoprotein Cholesterol; Apo-A, Apolipoprotein A; Apo-B, Apolipoprotein B; APB: APA, Apolipoprotein B to Apolipoprotein A Ratio; BUN, Blood Urea Nitrogen; CREA, Creatinine; UA, Uric Acid; GLU, Glucose

Table 2.

Clinical characteristics of the training and validation cohorts

Characteristic	Overall, N = 2421	test, N = 731	train, N = 1691	p-value2
label	140 (58%)	38 (52%)	102 (60%)	0.230
age	57.71 (10.53)	58.03 (10.30)	57.58 (10.66)	0.808
gender	168 (69%)	45 (62%)	123 (73%)	0.084
CHO	5.17 (1.89)	5.09 (1.79)	5.21 (1.93)	0.921
TG	2.42 (2.93)	2.38 (1.50)	2.44 (3.36)	0.070
HDL-C	1.06 (0.36)	1.05 (0.33)	1.06 (0.38)	0.839
LDL-C	3.14 (1.50)	3.05 (1.48)	3.18 (1.51)	0.659
Apo-A	1.21 (0.33)	1.20 (0.30)	1.22 (0.35)	0.983
Apo-B	1.12 (0.44)	1.12 (0.44)	1.12 (0.44)	0.801
APB: APA	0.97 (0.44)	0.99 (0.54)	0.96 (0.39)	0.999
BUN	12.56 (8.20)	12.54 (7.93)	12.56 (8.33)	0.947
CREA	294.41 (269.28)	279.72 (234.97)	300.76 (283.24)	0.609
BUN/CREA	0.06 (0.03)	0.06 (0.03)	0.06 (0.02)	0.345
UA	417.91 (114.84)	428.62 (123.83)	413.29 (110.79)	0.382
GLU	7.66 (3.50)	7.80 (3.92)	7.61 (3.32)	0.749

¹n (%); Mean (SD)

²Pearson’s Chi-squared test; Wilcoxon rank sum test

Abbreviations: DN, Diabetic Nephropathy; CHO, Total Cholesterol; TG, Triglycerides; HDL-C, High-Density Lipoprotein Cholesterol; LDL-C, Low-Density Lipoprotein Cholesterol; Apo-A, Apolipoprotein A; Apo-B, Apolipoprotein B; APB: APA, Apolipoprotein B to Apolipoprotein A Ratio; BUN, Blood Urea Nitrogen; CREA, Creatinine; UA, Uric Acid; GLU, Glucose

Fig. 2.

Workflow of the radiomics analysis. The process includes: Image acquisition and ROI segmentation of the whole kidney on two‑dimensional ultrasound images; Extraction of radiomics features (first‑order, shape, and texture features); Feature selection using LASSO regression; Model construction combining radiomics and clinical biochemical data; Model evaluation with ROC curves, calibration curves, and decision curve analysis

Feature selection and model development

A total of 1,025 radiomics features were extracted from each two‑dimensional ultrasound image. After inter‑observer reproducibility analysis (ICC > 0.75), 64 features (6.2%) with low reproducibility were excluded, and the remaining 961 features (93.8%) were retained for further analysis. LASSO regression was then applied to select the most predictive features for each classification task.

For the early vs. mid‑to‑late DN task, 21 features with non‑zero coefficients were selected at the optimal λ value of 0.032 (log λ = − 3.41). For the early‑to‑middle vs. late DN task, 10 features were selected at λ = 0.045 (log λ = − 3.08). The LASSO coefficient paths and cross‑validation error plots are shown in Fig. 3.

Fig. 3.

LASSO regression feature selection. (A) Coefficient path plot for the early vs. mid-to-late DN model; the vertical line indicates the optimal λ selected by 10‑fold cross‑validation. (B) Cross‑validation error curve for the early vs. mid-to-late DN model, showing the mean squared error as a function of log λ. (C) Coefficient path plot for the early-to-middle vs. late DN model. (D) Cross‑validation error curve for the early-to-middle vs. late DN model. The optimal λ values correspond to 0.032 (log λ = -3.41) and 0.045 (log λ = -3.08), respectivel

Diagnostic performance of the models

The diagnostic performance of different models for DN staging is presented in Table 3. For the early vs. mid‑to‑late DN task, the combined model (clinical features + radiomics) achieved an AUC of 0.939 (95% CI: 0.905–0.969) in the training cohort and 0.876 (95% CI: 0.787–0.950) in the validation cohort, with corresponding sensitivities of 0.882 and 0.816, and specificities of 0.896 and 0.686. DeLong’s test showed that the combined model significantly outperformed both the clinical model (p = 0.001) and the radiomics model (p < 0.001) in the training cohort. In the validation cohort, the combined model was significantly superior to the radiomics model (p = 0.021), while the difference with the clinical model was not statistically significant (p = 0.578).

Table 3.

Diagnostic performance of different models for DN staging

Task	Model	Cohort	AUC (95% CI)	Sensitivity	Specificity	F1 Score	p-value*
Early vs. Mid-to-Late DN	Clinical	Training	0.856 (0.798–0.907)	0.775	0.731	0.794	0.001
	Radiomics	Training	0.815 (0.742–0.874)	0.755	0.672	0.766	< 0.001
	Combined	Training	0.939 (0.905–0.969)	0.882	0.896	0.905	Ref.
	Clinical	Validation	0.896 (0.814–0.966)	0.842	0.8	0.831	0.578
	Radiomics	Validation	0.759 (0.632–0.874)	0.763	0.6	0.716	0.021
	Combined	Validation	0.876 (0.787–0.950)	0.816	0.686	0.775	Ref.
Early-to-Middle vs. Late DN	Clinical	Training	0.910 (0.856–0.954)	0.683	0.917	0.746	0.042
	Radiomics	Training	0.759 (0.683–0.831)	0.703	0.450	0.862	< 0.001
	Combined	Training	0.951 (0.920–0.975)	0.767	0.917	0.800	Ref.
	Clinical	Validation	0.924 (0.851–0.983)	0.852	0.913	0.852	0.15
	Radiomics	Validation	0.831 (0.711–0.925)	0.593	0.913	0.681	0.02
	Combined	Validation	0.955 (0.902–0.991)	0.889	0.913	0.873	Ref.

p-value from DeLong’s test comparing each model to the combined model within the same cohort and task. Ref. indicates the reference model for comparison

For the early‑to‑middle vs. late DN task, the combined model achieved an AUC of 0.951 (95% CI: 0.920–0.975) in the training cohort and 0.955 (95% CI: 0.902–0.991) in the validation cohort, with sensitivities of 0.767 and 0.889, and specificities of 0.917 and 0.913, respectively. DeLong’s test indicated that the combined model significantly outperformed the radiomics model in both training (p < 0.001) and validation (p = 0.02) cohorts, while the improvement over the clinical model was significant only in the training cohort (p = 0.042). The nomograms, ROC curves, calibration curves, and decision curves for both tasks are presented in Supplementary Figs. 1 and 2, demonstrating good calibration and positive net clinical benefit.

Model performance after excluding renal function markers

To assess the independent contribution of radiomics beyond traditional renal function indicators, we repeated the model‑building process after excluding BUN, CREA, and BUN/CREA from the clinical features. The results are summarized in Table 4. For the early vs. mid‑to‑late DN task, the combined model (clinical features without renal markers + radiomics) achieved an AUC of 0.860 (95% CI: 0.801–0.912) in the training cohort, with a sensitivity of 0.990 and specificity of 0.134. In the validation cohort, the AUC was 0.750 (95% CI: 0.623–0.871), with a sensitivity of 0.921 and specificity of 0.000. For the early‑to‑middle vs. late DN task, the combined model achieved an AUC of 0.670 (95% CI: 0.603–0.777) in the training cohort, with a sensitivity of 0.383 and specificity of 0.872. In the validation cohort, the AUC was 0.759 (95% CI: 0.635–0.865), with a sensitivity of 0.333 and specificity of 0.913.

Table 4.

Diagnostic performance of different models for DN staging

Task	Model	Cohort	AUC (95% CI)	Sensitivity	Specificity	F1 Score
Early vs. Mid-to-Late DN	Combined	Training	0.860 (0.801–0.912)	0.99	0.134	0.774
	Combined	Validation	0.750 (0.623–0.871)	0.921	0	0.648
Early-to-Middle vs. Late DN	Combined	Training	0.670 (0.603–0.777)	0.383	0.872	0.474
	Combined	Validation	0.759 (0.635–0.865)	0.333	0.913	0.450

p-value from DeLong’s test comparing each model to the combined model within the same cohort and task. Ref. indicates the reference model for comparison

These findings confirm that while radiomics provides complementary information, renal function markers remain essential for optimal DN staging. The substantial performance reduction—particularly in the early‑to‑middle vs. late DN task—underscores the central role of BUN and CREA as core indicators of disease progression. The near‑zero specificity observed in the validation cohort of the early vs. mid‑to‑late task warrants cautious interpretation and may reflect model instability after removal of key predictors.

Discussion

Accurate staging of diabetic nephropathy (DN) is crucial for guiding treatment strategies and predicting long-term prognosis. In recent years, ultrasound-based radiomics and machine learning have shown promise in the diagnosis and prognostication of kidney diseases [18–20]. In this study, we developed and validated nomograms integrating two-dimensional ultrasound radiomics and clinical biochemical data for DN staging, achieving excellent diagnostic performance across two clinically relevant binary classification tasks.

Key findings and the central role of renal function markers

Our results demonstrate that the combined models significantly outperformed models based solely on clinical or radiomics features in both tasks. Notably, renal function markers—particularly BUN and CREA—were consistently selected as important predictors. This finding aligns with established knowledge that these markers directly reflect glomerular filtration function and are intrinsically linked to DN progression [21–23]. Previous machine learning studies have similarly identified CREA and BUN as fundamental features for predicting DN [24, 25], underscoring their central role in disease assessment.

To address concerns about circular reasoning, we performed a sensitivity analysis excluding BUN, CREA, and BUN/CREA (Table 4). As expected, model performance dropped substantially across tasks. For the early‑to‑middle vs. late DN task, the validation AUC fell from 0.955 to 0.759, confirming that these markers are core indicators of disease progression. However, the early vs. mid‑to‑late DN task requires cautious interpretation. In the validation cohort, the combined model (without renal markers) achieved an AUC of 0.750 but with a specificity of 0.000 — meaning every patient was classified as “mid‑to‑late”. This occurred because all mid‑to‑late patients were correctly identified (sensitivity = 0.921), while no early‑stage patient was correctly recognized. Two factors may explain this result. First, class imbalance: the validation cohort contained only a few early‑stage patients. After removing BUN and CREA—the strongest discriminators—the model became biased toward the majority class (mid‑to‑late). Second, model instability: the modest sample size (n = 73) amplified this bias, causing the default probability threshold (0.5) to fall below all predicted probabilities for the majority class.

What does specificity = 0.000 imply for radiomics? It indicates that for distinguishing early from mid‑to‑late DN, radiomics features alone cannot compensate for the loss of renal function markers. While radiomics provides complementary information when combined with clinical data, it is not a substitute for BUN and CREA. Therefore, we qualify our earlier conclusion: the complementary value of radiomics is primarily observed in the early‑to‑middle vs. late DN task and in models that include renal markers; it does not imply that radiomics can replace renal function tests.

We explicitly acknowledge this as a fundamental limitation of our sensitivity analysis. The near‑zero specificity shows that model performance heavily depends on BUN and CREA, and that radiomics lacks sufficient independent signal for this specific task. Future studies with larger, balanced cohorts and optimized thresholds are needed to assess whether radiomics offers independent value for early vs. mid‑to‑late DN staging.

Complementary value of radiomics

Despite the performance reduction, the combined model retained some predictive ability after excluding renal markers, particularly in the early vs. mid-to-late DN task (validation AUC = 0.750). This suggests that radiomics captures complementary microstructural information not fully reflected by routine biochemical tests. In the early vs. mid-to-late DN task, the clinical and radiomics models individually showed weaker diagnostic performance compared to their combination. This finding supports the notion that traditional markers such as albuminuria and eGFR may lag behind actual pathological changes in early DN due to renal compensatory mechanisms [26]. Radiomics, by extracting high-dimensional texture features from ultrasound images, can potentially detect subtle tissue alterations that precede or accompany functional decline [27]. In the early-to-middle vs. late DN task, the combined model again demonstrated superior diagnostic performance, with higher sensitivity and specificity than either individual model. This further confirms that radiomics complements the limitations of clinical biochemical data, and the two modalities together provide a more comprehensive assessment of disease severity.

Comparison with previous studies

Our findings are consistent with and extend prior work in this field. Similar to Liu et al. [24] and Gao et al. [27], we identified CREA, BUN, and HDL-C as important predictors in machine learning models for DN. However, unlike previous studies that focused on DN prediction or detection [28–31], our study specifically addresses staging—a clinically more challenging task with direct implications for treatment decisions. By constructing two binary classification models targeting key clinical decision points (early vs. progressive disease, and non-advanced vs. advanced disease), we provide practical tools that align with real-world clinical reasoning.

Clinical implications

Our study has several important clinical implications. First, the strong performance of models including renal function markers reinforces that BUN and CREA remain indispensable for accurate DN staging in routine practice. Second, the residual predictive value of radiomics after excluding these markers suggests that ultrasound-based texture analysis can provide additional information about renal tissue microstructure, potentially offering insights into pathological changes that are not yet reflected by functional tests. However, the near-zero specificity observed in the sensitivity analysis cautions against relying solely on radiomics when renal function data are unavailable. Third, the integration of radiomics with readily available clinical data—both obtained during routine patient workup—makes our approach highly feasible for clinical translation without additional burden on patients or healthcare systems.

Limitations

This study has several limitations. First, its retrospective, single-center design may introduce selection bias and limits generalizability. The lack of external validation necessitates cautious interpretation of our results. Second, as a retrospective study, we were unable to fully standardize ultrasound image acquisition parameters across all patients, which may introduce variability in radiomics features. Third, the sample size, particularly in the validation cohort, is modest; this likely contributed to the wide confidence intervals and the unstable specificity observed in the sensitivity analysis. Fourth, our sensitivity analysis revealed that model performance declines substantially when renal function markers are excluded, indicating that radiomics alone cannot replace traditional biochemical tests for DN staging. Finally, we did not explore advanced imaging techniques such as contrast-enhanced ultrasound or elastography, which may provide additional complementary information.

Future directions

Prospective, multicenter studies with standardized imaging protocols are warranted to validate our models and evaluate their generalizability. Future work should also explore the integration of ultrasound radiomics with other modalities (e.g., shear wave elastography, contrast-enhanced ultrasound) and deep learning approaches to further enhance diagnostic accuracy. Additionally, longitudinal studies are needed to assess whether the radiomics signature can predict DN progression and treatment response, thereby supporting personalized patient management.

Conclusion

This study demonstrates that integrating two‑dimensional ultrasound radiomics with clinical biochemical data enables effective non‑invasive staging of diabetic nephropathy. Renal function markers (BUN, CREA) remain essential for staging accuracy, particularly for distinguishing early from mid‑to‑late DN. Radiomics provides complementary microstructural information that enhances model performance when combined with clinical data, especially for the early‑to‑middle vs. late DN task. However, the sensitivity analysis revealed that radiomics alone cannot replace renal function markers in the early vs. mid‑to‑late classification, highlighting a key limitation of the current approach. Future work should focus on multi‑center validation and methods to mitigate class‑imbalance issues.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Author contributions

Xuee Su and Baoyuan Xie conceived and designed the study. ShanHu Wu and CuiLiu Lin collected the data. Xuee Su and Shanhu Wu drafted the manuscript. Huaigang Wang and Jing Liu analyzed the data. Chengbao Peng and Hefan He critically revised the manuscript. All authors contributed to this article, approved the submitted version, and agreed to be accountable for all aspects of the work.

Funding

This work was supported by Joint funds for the innovation of science and technology, Fujian province (Grant number: 2023Y9244)and the Fujian Provincial Clinical Key Specialty Construction Project (No. HLZDZK202307).

Data availability

The primary data underpinning the findings of this study can be obtained by contacting the corresponding author directly.

Declarations

Ethics approval and consent to participate

The study adhered to the principles of the Declaration of Helsinki and received approval from the Medical Ethics Committee of the Second Affiliated Hospital of Fujian Medical University (approval No. 2023252). As this study was retrospective, the need for written informed consent was waived.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.