A Deep Learning Framework Based on Ultrasomics for IgA Nephropathy Detection: A Multi-Centre, Multi-Scanner Validation Study

A Deep Learning Framework Based on Ultrasomics for IgA Nephropathy Detection: A Multi-Centre, Multi-Scanner Validation Study | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article A Deep Learning Framework Based on Ultrasomics for IgA Nephropathy Detection: A Multi-Centre, Multi-Scanner Validation Study Dong Wang, Yanyan Wu, JinQi Gong, ZiHan Tang, MingRui Fu, Yunhao Luo, and 5 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8425698/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Immunoglobulin A nephropathy (IgAN), the most prevalent primary glomerulonephritis globally, lacks reliable non-invasive diagnostics. Current diagnosis depends on invasive renal biopsy (bleeding/pain risks), and conventional ultrasound misses early microstructural lesions. Here, we present IANet, an interpretable deep learning framework using Bilateral Renal Symmetry-Aware Analysis (BRSA) to amplify subtle IgAN-related sonographic changes. In a 4-centre retrospective study of 2,012 participants, IANet was trained (n = 1,256), internally (n = 538) and externally (n = 218) validated. It achieved 96.7% accuracy (95% CI: 95.8–97.5%) and 0.975 AUC (95% CI: 0.962–0.984) internally, with 80.1% lower false positive rate (3.1% vs. 15.6% for ResNet-18). Externally (six ultrasound manufacturers), it maintained 86.2% accuracy (95% CI: 81.2–90.4%), 0.928 AUC (95% CI: 0.891–0.955%), 96.2% sensitivity, and 77.0% specificity. Blinded tests vs. 3 radiologists (10–15 years’ experience) showed IANet outperformed humans (98.0% vs. 71.0–82.0% accuracy, p 1,000-fold faster (4 s vs. 28–36 min per 100 cases). Subgroup analysis stratified by the Oxford MEST-C classification demonstrated that IANet achieved its peak diagnostic accuracy for S1 and T2 lesions in IgAN, while a modest performance decline was observed in cases with T0 pathology. Histopathological validation confirmed that IANet's attention heterogeneity was significantly associated with chronic lesions, demonstrating strong predictive value for segmental sclerosis (S1, OR = 4.28, p < 0.001) and interstitial fibrosis/tubular atrophy (T1&T2, OR = 3.20, p < 0.001). In a DKD cohort (n = 192), it correctly classified 64.1% as non-IgAN. IANet is a novel tool for non-invasive and dynamic renal function monitoring, thereby facilitating the advancement of precision nephrology. Health sciences/Nephrology/Kidney diseases/Nephritis Health sciences/Diseases/Kidney diseases/Chronic kidney disease IgA Nephropathy IANet Bilateral Renal Symmetry-Aware Analysis (BRSA) Ultrasound Diagnosis Explainable AI Histopathological Correlation Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction Immunoglobulin A nephropathy (IgAN), characterized by dominant glomerular IgA deposition, is the leading cause of primary glomerulonephritis globally and accounts for 45% of such diagnoses in China 1 , 2 . Despite therapeutic advances (e.g., immunosuppressants and renin-angiotensin system inhibitors), > 30% of patients progress to end-stage kidney disease (ESKD) within 10–20 years of diagnosis 3 , 4 , imposing substantial clinical and economic burdens (e.g., > $ 2.4 billion annual ESKD management costs in the US alone 3 , 5 . The gold standard for IgAN diagnosis—renal biopsy with the Oxford MEST-C score 6 , 7 , 8 , 9 , is limited by three key barriers: (1) invasiveness, with a 2–5% risk of major complications (e.g., perirenal hematoma) 10 ; (2) variable regional biopsy thresholds 11 , leading to missed early intervention opportunities; (3) lack of integration with imaging biomarkers, as the MEST-C score relies solely on histology and clinical metrics. Renal ultrasound, the preferred non-invasive modality for chronic kidney disease (CKD), detects parenchymal echogenicity increases linked to renal fibrosis—a stronger predictor of IgAN outcomes than traditional biomarkers (e.g., proteinuria) 12 . However, this clinician-dependent approach introduces substantial inter-observer variability, compromising diagnostic reliability 12 , 13 (inter-observer agreement: Fleiss’ κ = 0.366 in our study) and insensitivity to early lesions 14 . Artificial intelligence (AI) has emerged to enhance ultrasound diagnostics 15 , 16 , but three critical barriers hinder clinical translation: (1) poor generalizability across heterogeneous scanners and acquisition protocols 17 , 18 , as most models are trained on single-centre, single-device datasets; (2) scarce histopathologically annotated data 15 , 19 , introducing population bias (e.g., overfitting to Asian cohorts) 10 ; (3) inability to detect subtle IgAN-specific microstructural changes amid anatomical noise from age-related or comorbidity-related alterations 20 , 21 . Prior AI models for IgAN either rely on manual region-of-interest (ROI) delineation 17 , 22 , which is a time-consuming step incompatible with clinical workflows, or focus on single-kidney feature extraction 15 without leveraging the physiological symmetry of bilateral kidneys, and this symmetry serves as a key insight for amplifying pathological differences. To address these gaps, we developed IANet , a deep learning framework that: (1) uses BRSA to enhance detection of early, asymmetric lesions; (2) enables fully automated kidney localization (eliminating manual ROI bias); (3) ensures interpretability via Grad-CAM (validated against histopathology and biomarkers, Table 5 ); (4) maintains robustness across multi-centre, multi-scanner datasets. Our results demonstrate IANet is a potential non-invasive tool to provide imaging support for IgAN disease course Results 1. Study Population and Baseline Characteristics Figure 1 outlines the study flowchart A total of 2,012 participants (four institutions in China) were enrolled: 963 biopsy-confirmed IgAN patients (Oxford MEST-C stages: M0/M1=245/511, E0/E1=623/131, S0/S1=589/167, T0/T1/T2=328/287/131, C0/C1-C2=652/104) and 1,049 healthy controls (normal renal function, no hematuria/hypertension). The cohort included 1,077 males (53.53%) and 935 females (46.47%), with a mean age of 34.07 ± 12.05 years (range: 18–78 years). Data from the primary centre (Sichuan Provincial People’s Hospital, Chengdu, n=1,794) were split into training (n=1,256: 601 IgAN, 655 controls) and internal test sets (n=538: 257 IgAN, 281 controls) at a 7:3 ratio using stratified random sampling (preserving IgAN stage distribution). As shown in Figure 2, the development process of IANet encompasses data curation, model training (Figure 2a), and multi-stage evaluation (Figure 2b). The external validation cohort (n=218: 105 IgAN, 113 controls) included participants from three independent centres: The First Affiliated Hospital of Chongqing Medical University (Chongqing), Sichuan Integrative Medicine Hospital (Chengdu), and The First Affiliated Hospital of Ningbo University (Ningbo). All data were analyzed using ultrasound images of the maximal longitudinal sections of bilateral kidneys (Figure 3). Baseline demographic and clinical characteristics (Table 1) showed no significant differences in age or sex between the IgA nephropathy (IgAN) group and the control group across the training, internal test, and external validation cohorts. However, significant differences were observed in key clinical biomarkers (all p < 0.001). In all three cohorts, serum creatinine levels were higher in the IgAN group compared to the control group, while the eGFR was lower in the IgAN group. These findings are consistent with the known pathophysiological characteristics of IgA nephropathy. To ensure full transparency regarding the imaging data sources, the sample distribution across all participating centres and ultrasound manufacturers is detailed in Supplementary Table 1. This distribution reflects real-world clinical usage patterns. Briefly, the four major manufacturers (Mindray, Philips, Samsung, and GE) collectively contributed 84.2% of the total dataset (1,695/2,012 samples), with Mindray being the single most prevalent brand (28.2%). 2. Model Performance: Baseline vs. IANet IANet outperformed four baseline architectures (ResNet-18, ResNet-50, DenseNet-121, InceptionV3) across all metrics (all p<0.001; Table 2). In internal validation: ResNet-18-based IANet (prioritized for clinical translation due to balance of performance and efficiency) achieved 96.73% accuracy (95% CI: 95.81–97.49%) and 0.975 AUC (95% CI: 0.962–0.984)—12.26% higher accuracy and 0.083 higher AUC than baseline ResNet-18 (84.47% accuracy, 0.892 AUC). False positive rate (FPR) was reduced by 76.80% vs. baseline ResNet-18 (3.1% [95% CI: 2.3–4.0%] vs. 15.6% [95% CI: 13.2–17.9%]), minimizing unnecessary follow-ups. Positive predictive value (PPV) = 96.8% (95% CI: 94.1–98.7%) and negative predictive value (NPV) = 95.8% (95% CI: 92.6–97.9%), confirming reliability in both IgAN identification (high PPV) and rule-out (high NPV). ResNet-50/DenseNet-121/InceptionV3-based IANet variants also showed significant improvements over their baselines (e.g., ResNet-50: 88.62% → 95.82% accuracy, p<0.001), but ResNet-18-based IANet had the lowest computational cost (inference time: 0.04 s per case vs. 0.12 s for ResNet-50-based IANet) and highest generalizability (external validation AUC: 0.9275 vs. 0.901 for ResNet-50-based IANet; Supplementary Table 2). 1. Innovation Over Existing Research IANet addresses three unmet needs in IgAN diagnostics, distinguishing it from prior AI models: BRSA for asymmetry amplification : Unlike single-kidney models 15, 22 , BRSA amplifies pathological asymmetries (e.g., subtle cortical echogenicity differences between left/right kidneys) while suppressing symmetric anatomical noise (e.g., age-related parenchymal changes). Ablation studies (Table 2) confirm BRSA drives a 12.26% accuracy gain in ResNet-18 variants—critical for detecting early T0-stage lesions. Fully automated localization : Eliminates manual ROI delineation 23 , which is a bottleneck in prior studies 17 . Grad-CAM and Table 5 confirm IANet accurately identifies kidneys in complex abdominal scans (e.g., amid liver/spleen) and focuses on pathologically relevant regions. This automation reduces inter-operator variability and enables real-time clinical integration. Multi-modal interpretability : Unlike “black-box” AI 24, 25 , IANet provides interpretability by linking its attention heterogeneity directly to underlying pathology. This heterogeneity, quantified as GLCM-ASM was significantly associated with chronic fibrotic lesions (e.g., S1, T1&T2) 26 , demonstrating that the model detects IgAN by capturing parenchymal structural disarray, thereby building crucial clinical trust. Compared to published IgAN AI models (Supplementary Table 6), IANet achieves higher external validation AUC (0.9275 vs. 0.82–0.89) and faster processing (0.04 s vs. 1–5 s per case), with the unique advantage of bilateral analysis. 3. IANet Performance in External Validation In external validation (Table 3), IANet maintained robust performance across heterogeneous scanners (6 manufacturers: GE, Philips, Samsung, Mindray, Toshiba, Canon) and centres: Accuracy = 86.24% (95% CI: 81.15–90.37%), AUC = 0.9275 (95% CI: 0.8912–0.9548%); Sensitivity = 96.19% (95% CI: 91.57–98.64%) (minimizing missed diagnoses), specificity = 77.00% (95% CI: 68.53–83.98%); Confusion matrix: True Positives (TP)=101, True Negatives (TN)=87, False Positives (FP)=26, False Negatives (FN)=4; Processing time for 218 samples = 8.85 s (0.04 s per case), with a false negative rate of 1.83% (4/105 IgAN cases misclassified, all T0 stage with IFTA <5%). Cross-device performance (Supplementary Table 2) confirmed no significant differences in accuracy across major manufacturers (F=1.23, p=0.30). The lower overall specificity (77.00%) vs. individual device specificity (92.9–94.1%) was attributed to 26 false positives from two low-volume manufacturers, which were included to reflect real-world device diversity. Prior multi-center studies have reported significant performance drops (5-12% accuracy reduction) when models are tested across heterogeneous scanner types 17 . 4. Ablation Study: Impact of Bilateral Interaction Ablation experiments compared full IANet (with BRSA modules: cross-interaction, enhanced difference, multi-dimensional fusion) to baseline models (concatenated bilateral inputs without BRSA). The full model showed significant improvements (Table 2): ResNet-18-based IANet: Accuracy 84.47% → 96.73% (+12.26%, 95% CI: 10.12–14.18%), AUC 0.892 → 0.975 (+0.083), FPR 15.6% → 3.1% (80% reduction); ResNet-50-based IANet: FPR 11.2% → 4.5% (59.8% reduction), sensitivity 85.7% → 93.8% (+8.1%); These results confirm that cross-interaction and enhanced difference modules (core of BRSA) drive performance gains by amplifying pathological asymmetries (e.g., subtle cortical echogenicity differences between left/right kidneys) while suppressing symmetric anatomical noise. 5. IANet vs. Radiologists In a blinded assessment of 100 cases (42 IgAN, 58 controls; image quality score ≥4/5, excluding motion artifacts), Radiologists were provided only with the de-identified ultrasound images and were blinded to the patients' clinical history, laboratory results (e.g., eGFR), and biopsy findings, to ensure a direct comparison of visual feature extraction capabilities against the IANet model. IANet outperformed three board-certified radiologists (10–15 years of nephrology ultrasound experience; Table 4): Accuracy: IANet 98.0% (95% CI: 95.3–100.0%) vs. Radiologist X.G (15-yr exp) 82.0% (95% CI: 74.4–89.6%), Radiologist J.S (10-yr exp) 74.0% (95% CI: 65.5–82.5%), Radiologist F.H (12-yr exp) 71.0% (95% CI: 62.2–79.8%) (all p<0.001, Cochran’s Q test + post-hoc McNemar with Bonferroni adjustment, α=0.0083, supplementary table 3); Predictive values: IANet PPV=100% (no false positives), NPV=96.7% vs. radiologists PPV=68.9–79.4%, NPV=67.9–83.3%; Processing time: IANet=4 s vs. radiologists=28–36 min (>1,000-fold faster, critical for high-volume clinical settings). Consistent with the accuracy results, the receiver operating characteristic (ROC) curve analysis further validated IANet's superior discriminative capability. In the internal validation cohort, IANet (ResNet-18) achieved an AUC of 0.975 (95% CI: 0.962–0.984), which was significantly higher than that of the ResNet-18 baseline (0.892, 95% CI: 0.871–0.910) and ResNet-50 baseline (0.921, 95% CI: 0.903–0.936), as well as the reported AUC of conventional ultrasound (0.820, 95% CI: 0.79–0.85) (Supplementary Figure 1A). In the blinded 100-case assessment, IANet exhibited an outstanding AUC of 0.992 (95% CI: 0.978–1.000), substantially outperforming all three radiologists (Radiologist X.G: 0.864, 95% CI: 0.792–0.936; Radiologist J.S: 0.782, 95% CI: 0.700–0.864; Radiologist F.H: 0.751, 95% CI: 0.665–0.837) (Supplementary Figure 1B). Inter-rater agreement among radiologists was substantial (Fleiss’ κ=0.366, 95% CI: 0.253–0.480; p<0.001; supplementary table 4), highlighting subjective interpretation bias—an issue IANet eliminates via automated, objective analysis. IANet’s superior performance is supported by its specialized architecture designed for bilateral renal ultrasound analysis (Figure 4), which integrates bilateral symmetry-aware modules to capture pathological asymmetries critical for IgAN detection We applied gradient-weighted class activation mapping (Grad-CAM) to visualize spatial attention patterns in IANet’s decision-making process. As shown in Figure 5, heatmaps (red/orange = high attention; green/blue = low attention) reveal the model’s diagnostic rationale through region-specific feature activation. Consistent anatomical focus in the Grad-CAM heatmap confirms that IANet can automatically locate kidneys in ultrasound images containing complex anatomical structures such as liver, spleen, and subcutaneous fat, achieving precise spatial alignment with ultrasound-defined renal borders (Fig. 5). Meanwhile, IANet exhibits attention to adjacent organs (liver/spleen), indicating that this may be consistent with radiologists’ diagnostic thinking of contrasting the kidneys with surrounding structures. All diagnostic errors occurred exclusively in early Oxford T0-stage IgAN (interstitial fibrosis and tubular atrophy [IFTA] <25%): IANet missed 2 cases (2% error rate), both with minimal fibrosis (IFTA <5%; one with predominant podocyte injury, no detectable echogenicity changes); Radiologists missed 10–26 cases (17.2–44.8% error rates), primarily T0-stage cases with IFTA 5–15%; Eight cases were missed by all evaluators (including IANet’s 2 cases), all with IFTA <15%—indicating a shared diagnostic blind spot for ultra-early lesions with non-specific sonographic features. This observation that misclassifications were confined solely to early-stage disease prompted us to systematically evaluate the model’s performance across the full spectrum of Oxford MEST-C pathological phenotypes in both internal and external validation cohorts. With only 18 M1 cases (5.0% of the combined internal and external validation cohort, n=362 [257+105])) in the M subgroup; insufficient sample size limits statistical power for reliable subgroup analysis, thus, the M subgroup is excluded, with only E, S, T, C included for subsequent analyses. The results demonstrated that IANet maintained high accuracy for advanced IgAN stages but showed reduced performance for ultra-early disease (Supplementary Table 5): High-performance stages: S1 (97.0% [95% CI: 93.6–98.8%]), T1-T2 (94.8–96.2%). Low-performance stage: T0 (IFTA <25%: 83.6% [95% CI: 79.2–87.3%]), primarily due to minimal fibrosis (<5% IFTA) in 12 misclassified cases. 6. Correlation of IANet Attention with Histopathology Analysis of Grad-CAM attention maps revealed that IgA nephropathy (IgAN) cases with chronic lesions presented significantly higher attention distribution heterogeneity in the renal parenchyma. Based on this observation, we hypothesized that the degree of IANet’s attention heterogeneity correlates with lesion pathological subtypes, and verified this association using texture uniformity analysis based on gray level co-occurrence matrix-angular second moment (GLCM-ASM) (Table 5). Parenchymal attention heterogeneity showed strong, significant associations with chronic fibrosing pathologies. It served as a powerful independent predictor for segmental sclerosis (S1) (OR=4.28, 95% CI: 2.95–6.21, p<0.001, n=198, 54.7% of validation cohort) and interstitial fibrosis/tubular atrophy (T1&T2) (OR=3.20, 95% CI: 2.25–4.55, p<0.001, n=129, 35.6% of validation cohort). In contrast, its associations with acute lesions were markedly weaker: a significant but modest link was found with endothelial proliferation (E1) (OR=2.15, 95% CI: 1.32–3.51, p=0.002, n=97, 26.8% of validation cohort), while the association with crescent formation (C1&C2) showed a positive trend without reaching statistical significance (OR=1.83, 95% CI: 0.97–3.45, p=0.063, n=114, 31.5% of validation cohort). Mesangial proliferation (M1) was excluded from analysis due to insufficient sample size (n=18, 5.0%). As shown in Figure 5, renal parenchyma in healthy controls (Figure 5a) exhibited homogeneous attention distribution. Conversely, correctly classified IgAN cases (Figure 5b) presented high attention heterogeneity, with focus concentrated on regions of abnormal cortical echogenicity. A challenging T0-stage case (IFTA 10–15%) missed by all radiologists but identified by IANet (Figure 5c) demonstrated the model’s capacity to detect subtle, low-heterogeneity pathological cues. Meanwhile, the model’s misdiagnosis of an ultra-early T0-stage case (Figure 5d) was characterized by homogeneous attention, consistent with the absence of sonographically discernible fibrosis. A representative S1T2 case (Figure 5e) displayed patchy, heterogeneous attention patterns that matched the distribution of chronic pathological lesions. Additionally, cortical echogenicity elevation showed a strong linear correlation with fibrosis area ratio (r=0.69, 95% CI: 0.59–0.77, p<0.001, n=87) 13 . These results confirm that IANet’s diagnostic capability originates from its sensitivity to parenchymal structural disarray, enabling effective capture of chronic sclerotic and fibrotic damage while showing limited sensitivity to acute inflammatory changes. 7. Specificity Assessment Against Diabetic Kidney Disease To evaluate differential diagnostic capability, IANet was tested on a biopsy-proen DKD cohort (n=192, excluded from model training; mean age 56.2±9.8 years, 58% male, eGFR 68.3±15.2 mL/min/1.73m²): IANet correctly classified 123 DKD cases as non-IgAN (specificity=64.1%), 69 cases (35.9%) were misclassified as IgAN—attributed to shared sonographic features (e.g., cortical echogenicity increase due to fibrosis) between IgAN and DKD. Subgroup analysis showed misclassified DKD cases had higher fibrosis scores (IFTA ≥25%: 78% vs. 32% in correctly classified cases, p<0.001)—highlighting the need for multimodal refinement (e.g., integrating HbA1c or urinary biomarkers). The specificity of IANet for differentiating IgAN from DKD (64.1%) is comparable to or higher than previously reported ultrasound AI models 15, 27 but remains suboptimal. The 35.9% false positive rate is mainly due to shared sonographic features (e.g., cortical fibrosis-induced echogenicity increase) between the two diseases. Future refinement should integrate multimodal data (e.g., HbA1c, urinary biomarkers) to enhance differential diagnostic capability, as demonstrated by recent studies showing that combining ultrasound and clinical parameters can improve DKD-IgAN discrimination to 72.8% specificity 28 . Discussion 2. Clinical Translation Potential IANet’s design aligns with unmet clinical needs across diverse healthcare settings: Pre-biopsy screening : High NPV (96.7%) enables reliable IgAN rule-out, reducing unnecessary biopsies by 32% in simulated cohorts (consistent with internal validation data). For example, in patients with isolated microscopic hematuria (a common IgAN presentation), IANet could identify low-risk cases (probability <20%) that avoid biopsy—reducing complication risks and healthcare costs. Multi-centre standardization : Cross-device performance (Supplementary Table 2) confirms IANet works across mainstream ultrasound manufacturers (GE, Philips, Samsung, Mindray) with no significant accuracy differences (p=0.30). This standardization addresses the problem of variable ultrasound interpretation across centres—critical for multi-centre trials and global adoption. Early intervention enabler : The model’s high sensitivity (96.19%) in external validation ensures minimal missed diagnoses, while its ability to detect M1/S1 stages (accuracy >96%) enables early initiation of renoprotective therapy, which is crucial to reducing ESKD progression. 3. Limitations and Future Directions Ultra-early disease (T0 stage): The reduced accuracy in T0-stage IgAN (83.6%) is attributed to the inherent limitation of BRSA: ultra-early lesions (IFTA <25%) typically exhibit bilateral symmetry (e.g., uniform mild cortical echogenicity), making it difficult for the cross-interaction and enhanced difference modules to amplify pathological disparities. This is consistent with our histopathological findings that T0-stage cases have minimal structural disorganization (higher GLCM-ASM values, mean=0.0012 vs. 0.00078 in T1/T2 stages), resulting in insufficient discriminative features. Future work will integrate quantitative ultrasound elastography 13 (to measure tissue stiffness) and integrate relative clinical parameters to improve ultra-early detection 29 . DKD specificity : 35.9% false positives highlight overlapping sonographic features between IgAN and DKD (e.g., fibrosis). Multimodal fusion with genomic data 21 or clinical parameters 30 will enhance differential diagnosis. We are currently expanding our research to develop a predictive model that incorporates multimodal ultrasound 31 and key clinical parameters to improve diagnostic capability. Pathological spectrum bias: The retrospective design of our study inherently limited the number of available M1-grade cases (n=18, 5.0%), constraining the analysis of its correlation with IANet's attention. Consequently, our pathological validation primarily underscores the model's strength in identifying chronic, fibrotic changes (S1, T1&T2), while the association with active inflammatory lesions requires future prospective studies with targeted enrollment. Device-related variability : Lower specificity in Toshiba/Canon devices (69.2–71.4%, Supplementary Table 2) is attributed to differences in image resolution (512×512 vs. 1024×1024 for GE/Philips) and acquisition protocols (e.g., gain settings) 32, 33 . We are developing a device-specific calibration module (based on transfer learning) to standardize image features 34 , with initial tests on Toshiba devices showing 12.5% specificity improvement (from 71.4% to 83.9%, n=28 cases) 4. Conclusion IANet provides a non-invasive, interpretable, and scalable tool for IgAN detection. Its high performance (internal AUC=0.975, external AUC=0.9275), speed (0.04 s per case), and validation against histopathology/biomarkers make it a strong candidate for clinical adoption. By enabling early, standardized diagnosis, IANet has the potential to optimizing clinical management pathways for IgAN by integrating standardized imaging metrics into patient evaluation. Methods 1. Study Design and Participants This multi-centre retrospective case-control study included four institutions in China: Primary centre: Sichuan Provincial People’s Hospital (Chengdu) – model training (n=1,256) and internal validation (n=538). External centres: The First Affiliated Hospital of Chongqing Medical University (Chongqing), Sichuan Integrative Medicine Hospital (Chengdu), The First Affiliated Hospital of Ningbo University (Ningbo) – external validation (n=218 total). Inclusion criteria (IgAN group): ≥18 years old; biopsy-confirmed IgAN (per Oxford MEST-C classification 8 ); renal ultrasound performed within 30 days of biopsy (to ensure structural-pathological consistency); complete clinical data (eGFR, serum creatinine, proteinuria). Inclusion criteria (control group): Healthy volunteers from hospital health management centres; normal renal function (eGFR ≥90 mL/min/1.73m²); no abnormal urinalysis (no hematuria/proteinuria); no history of kidney disease, hypertension, or diabetes. Exclusion criteria: Incomplete clinical/imaging data; poor-quality ultrasound images (motion artifacts, insufficient kidney visualization); secondary IgA deposition (e.g., lupus nephritis, Henoch-Schönlein purpura); coexisting autoimmune diseases (e.g., rheumatoid arthritis). The study was approved by the Institutional Review Board of Sichuan Provincial People’s Hospital (No. 2025-294). Informed consent was waived due to the retrospective design, with all data de-identified per HIPAA guidelines. 2. Data Acquisition and Preprocessing Ultrasound protocol : Uniform acquisition using curvilinear array transducers (2–5 MHz) across 6 manufacturers (GE, Philips, Samsung, Mindray, Toshiba, Canon). For each participant, a board-certified sonographer acquired 3–5 maximal longitudinal sections of each kidney (left/right); the section with the clearest corticomedullary junction was selected (final dataset: 3,588 images; 1,794 participants × 2 kidneys). Image quality assessment : A radiologist (10-yr exp) scored images on a 5-point scale (1=uninterpretable, 5=excellent visualization); only images with scores ≥3 were included. Image Preprocessing: All preprocessing was implemented in Python 3.9 using OpenCV 4.8.0 and PyTorch 2.0.1. To standardize inputs across heterogeneous scanners, all images were resized to 512×512 pixels using bicubic interpolation. This resolution was chosen based on pre-experiments balancing three criteria: optimal retention of the corticomedullary junction features (Dice Similarity Coefficient = 0.91), computational efficiency (0.04 s inference time per image on an NVIDIA Tesla 4070 GPU), and compatibility with both high- and low-resolution native scanner outputs. Data Augmentation: To improve model generalizability, a series of augmentations were applied only to the training set, with all randomness controlled by a fixed seed (seed=42) for reproducibility. This included: (1) Addition of Gaussian noise (σ = 0.01) to simulate inherent ultrasound signal variability. (2) Random rotation within a range of ±15°. (3) A centered crop retaining 85% of the original area to focus the model on the renal parenchyma. ()4 Horizontal flipping with a 50% probability, applied strategically to kidney pairs to maintain anatomically meaningful bilateral symmetry for the BRSA module, with careful preservation of original side labels. 3. Model Architecture The IANet architecture integrates five synergistic modules (Figure 4, labeled signal flow), comprising: Shared Feature Extraction Module Backbone: ResNet-18 pretrained on ImageNet 35 —selected for its balance of performance and efficiency. Processing pipeline: 7×7 convolution (stride=2, 64 filters) → batch normalization 36 → ReLU activation → max-pooling (3×3, stride=2) → four residual blocks (layer1: 64 filters; layer2: 128 filters; layer3: 256 filters; layer4: 512 filters) → 512-channel feature maps (spatial resolution: 16×16, 1/32 of input). Individual Feature Branches Two symmetric, independent branches (one for left kidney, one for right kidney) to extract kidney-specific features. Each branch includes: Channel-wise attention (1×1 convolution + sigmoid activation) to recalibrate feature importance → 1×1 convolution to compress features from 512→256 channels (reducing computational overhead). Cross-Interaction Module Multi-head attention (4 heads) to model bidirectional feature dependencies between left and right kidneys. Processing steps were flatten 256×16×16 feature maps into 256×256 sequences. Compute attention weights for left→right and right→left interactions (modeling pathological asymmetries). Reconstruct spatial dimensions (256×16×16) to preserve anatomical localization. Enhanced Difference Module Three-stage pipeline to amplify discriminative inter-kidney disparities: Absolute differencing: Compute pixel-wise feature deviations (|f_left − f_right|) to capture baseline asymmetries. Convolutional enhancement: 3×3 convolution (512 filters) + batch normalization + GELU activation 37 to extract multi-scale differential patterns. Joint spatial-channel attention: Channel weighting (reduction MLP 38 , hidden dimension=128) + 7×7 convolution (spatial focusing) to amplify disease-specific biomarkers (e.g., cortical echogenicity differences). Multi-Dimensional Fusion and Classification Feature integration: (1) Global average pooling: Compress cross-interaction features (256×16×16) into 256-dimensional vectors; compress difference features (512×16×16) into 512-dimensional vectors. (2) Concatenation: Combine vectors into 768-dimensional unified representations (correction from prior 1024D, reflecting actual feature dimensions). (3) Projection: Layer normalization → SiLU Sigmoid-weighted linear units for neural network function approximation in reinforcement learning → dropout (rate=0.3) to prevent overfitting. Classification: Linear layer (1 output) + sigmoid activation to generate IgAN probability scores (0–1). 4. Training Protocol The model was trained with rigorous protocols to ensure reproducibility and performance. All modules were initialized with a fixed random seed (seed=42). The ResNet-18 backbone was initialized with weights pretrained on ImageNet, while the remaining components—including the convolutional layers in the Individual Feature Branches, Enhanced Difference module, and Spatial-Channel Attention module—were initialized using He normal initialization 35 . The multi-head attention layer within the Cross-Interaction module was initialized with Xavier uniform initialization, and all bias terms were set to zero. Hyperparameters were optimized via 5-fold cross-validation on the training cohort (n=1,256) using a grid search strategy, with validation loss as the primary selection metric. The comprehensive grid search results for key hyperparameters (optimizer, learning rate, batch size, etc.) are detailed in Supplementary Table 7. The final model was trained for 300 epochs using the Adam optimizer with a learning rate of 0.001, which provided superior convergence speed and lower validation loss compared to Stochastic Gradient Descent (SGD). A batch size of 32 was selected to maximize the memory utilization of the NVIDIA Tesla 4070 GPU. To mitigate overfitting, L2 weight decay (1×10⁻⁴) and dropout (rate=0.3) were applied. Gradient clipping (max norm=1.0) was employed to prevent gradient explosion. The loss function was binary cross-entropy, with class weights set to 1:1.08 to balance the IgAN and control classes 39 . Early stopping was triggered if the validation loss failed to improve for 15 consecutive epochs, and the model checkpoint with the lowest validation loss was retained. To quantitatively evaluate the contribution of the proposed Bilateral Renal Symmetry-Aware (BRSA) components, an ablation study was conducted on the internal validation cohort (n=538). Four model variants—all sharing the same ResNet-18 backbone and training hyperparameters—were compared: (1) a baseline model without any BRSA components; (2) a model with only the Cross-Interaction module; (3) a model with only the Enhanced Difference module; and (4) the full IANet incorporating all BRSA submodules. 5. Histopathological and Biomarker Correlation Pathologist review : Two renal pathologists (>15-yr exp) independently reviewed Grad-CAM heatmaps and corresponding biopsy slides in a blinded manner (no knowledge of model outputs). Correlation methods: To quantify the heterogeneity of IANet's attention, the renal parenchyma was first manually segmented on each ultrasound image by a radiologist (10-yr exp) blinded to the model and pathology results. This defined the parenchymal region of interest (ROI). Then, gray level co-occurrence matrix-angular second moment (GLCM-ASM) was adopted to characterize parenchymal texture uniformity (and indirectly reflect attention heterogeneity, with lower ASM indicating higher texture/attention heterogeneity). The GLCM-ASM calculation for irregular ROIs included mask preprocessing, valid pixel pair screening (4 directions, d=1 pixel), symmetric and normalized GLCM construction, and arithmetic averaging of 4-direction ASM values. The mesangial proliferation (M1) subgroup was excluded due to insufficient sample size (n=18, 5.0%). 6. Statistical Analysis All statistical analyses were performed using Python 3.9 (with SciPy 1.10.1, scikit-learn 1.2.2, and Statsmodels 0.13.5 packages) and SPSS 29.0, with visualization conducted using Matplotlib 3.7.1. Model performance was evaluated primarily by the area under the receiver operating characteristic curve (AUC). The 95% confidence intervals (CIs) for AUC values were computed using the DeLong method, while the Wilson score interval method was employed for proportions such as accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). The sample size for the internal validation cohort was calculated to be 803 participants to achieve 95% accuracy with a ±1.5% precision. After data quality filtering, 538 participants were ultimately included. For the external validation cohort, a sample size of 198 was calculated to be necessary to detect an AUC difference of 0.10 compared to conventional ultrasound (AUC=0.82), with a two-sided α of 0.05 and 90% power; 218 participants were enrolled, achieving a power of 92.3%. To mitigate potential class imbalance, class weights were set at 1:1.08 based on the prevalence of IgAN (47.9%) in the training cohort. For statistical comparisons, categorical variables were analyzed using the Chi-square test or Fisher's exact test when cell counts were less than five. Continuous variables that were normally distributed were compared using the independent samples t-test, while the Mann-Whitney U test was applied for non-normally distributed data. Comparisons of diagnostic performance among multiple radiologists were conducted using Cochran's Q test, followed by post-hoc McNemar tests with a Bonferroni-adjusted significance level (α = 0.0083). Pairwise comparisons of AUC values were performed using DeLong's test. Inter-rater agreement was assessed using Fleiss' κ, interpreted as follows: 0.6, substantial. Declarations Author Contributions Yanyan Wu: Conceptualization, methodology, writing—original draft. Jinqi Gong: Software, model training, writing—original draft. MingRui Fu: Formal analysis, data curation. ZiHan Tang: Visualization. Yunhao Luo, XueYing Chen, Yong Jin: Clinical data collection (ultrasound images, biopsy reports). Chengyuan Ye: Data curation, project administration. XinYi Hu: Formal analysis, validation. Dong Wang, Ping Zhang: Supervision, writing—review & editing (validated statistical analyses). All authors discussed the results and contributed to the final manuscript. Yanyan Wu and Jinqi Gong contributed equally to this work as co-first authors. Dong Wang and Ping Zhang jointly supervised the project and served as co-corresponding authors. Code availability The code used to train and evaluate the model is available on GitHub (https://github.com/wyydoctor/IgAN_model.git). Competing Interests The authors declare no competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Acknowledgments We thank Professor Guisen Li (Department of Nephrology, Sichuan Provincial People’s Hospital) for clinical insights into IgAN diagnosis and management, and the three renal pathologists (Drs. Li, Wang, and Zhang) for blinded histopathological review of biopsy slides (critical for Table 5 analyses). Ethics Statement The protocol of this study has been approved by the Institutional Review Board of Sichuan Provincial People’s Hospital (No. 2025-294). Due to the retrospective nature of the study, it meets the requirements for waiver of informed consent. We confirm that we have strictly adhered to the original protocol for data processing authorized by the IRB Funding This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. References Li LS, Liu ZH. Epidemiologic data of renal diseases from a single unit in China: analysis based on 13,519 renal biopsies. Kidney Int 66, 920–923 (2004). Filippone EJ, Gulati R, Farber JL. Contemporary review of IgA nephropathy. Front Immunol 15, 1436923 (2024). Pattrapornpisut P, Avila-Casado C, Reich HN. IgA Nephropathy: Core Curriculum 2021. Am J Kidney Dis 78, 429–441 (2021). Wyatt RJ, Julian BA. IgA nephropathy. N Engl J Med 368, 2402–2414 (2013). Cheng HT, Xu X, Lim PS, Hung KY. Worldwide Epidemiology of Diabetes-Related End-Stage Renal Disease, 2000–2015. Diabetes Care 44, 89–97 (2021). Barbour SJ, et al. Evaluating a New International Risk-Prediction Tool in IgA Nephropathy. JAMA Intern Med 179, 942–952 (2019). Suzuki H, et al. Aberrantly glycosylated IgA1 in IgA nephropathy patients is recognized by IgG antibodies with restricted heterogeneity. J Clin Invest 119, 1668–1677 (2009). Roberts IS, et al. The Oxford classification of IgA nephropathy: pathology definitions, correlations, and reproducibility. Kidney Int 76, 546–556 (2009). Trimarchi H, et al. Oxford Classification of IgA nephropathy 2016: an update from the IgA Nephropathy Classification Working Group. Kidney Int 91, 1014–1021 (2017). Hogan JJ, Mocanu M, Berns JS. The Native Kidney Biopsy: Update and Evidence for Best Practice. Clin J Am Soc Nephrol 11, 354–362 (2016). Kamano C, et al. A cross-sectional study in patients with IgA nephropathy of correlations between clinical data and pathological findings at the time of renal biopsy: a Japanese prospective cohort study. Clin Exp Nephrol 25, 509–521 (2021). Moghazi S, et al. Correlation of renal histopathology with sonographic findings. Kidney Int 67, 1515–1520 (2005). Libório AB, de Oliveira Neves FM, Torres de Melo CB, Leite TT, de Almeida Leitão R. Quantitative Renal Echogenicity as a Tool for Diagnosis of Advanced Chronic Kidney Disease in Patients with Glomerulopathies and no Liver Disease. Kidney Blood Press Res 42, 708–716 (2017). Chen Z, et al. Development and Deployment of a Novel Diagnostic Tool Based on Conventional Ultrasound for Fibrosis Assessment in Chronic Kidney Disease. Acad Radiol 30 Suppl 1, S295-s304 (2023). Qin X, Xia L, Ma Q, Cheng D, Zhang C. Development of a novel combined nomogram model integrating deep learning radiomics to diagnose IgA nephropathy clinically. Ren Fail 45, 2271104 (2023). Feretzakis G, et al. Emerging Trends in AI and Radiomics for Bladder, Kidney, and Prostate Cancer: A Critical Review. Cancers (Basel) 16, (2024). Peng T, et al. A multi-center study of ultrasound images using a fully automated segmentation architecture. Pattern Recognition 145, (2024). Lu WW, Zhang D, Ni XJ. A Review of the Role of Ultrasound Radiomics and Its Application and Limitations in the Investigation of Thyroid Disease. Med Sci Monit 28, e937738 (2022). Lundberg SM, Lee SI. A Unified Approach to Interpreting Model Predictions. In: 31st Annual Conference on Neural Information Processing Systems (NIPS)) (2017). Huang X, Hu Y, Zhang Y, Zhou Q. Two-Dimensional Ultrasound-Based Radiomics Nomogram for Diabetic Kidney Disease: A Pilot Study. Int J Gen Med 17, 1877–1885 (2024). Hasegawa K, et al. Inflammation-associated molecules in the glomerular-endothelium in mild IgA-nephropathy patients identified by single-cell and spatial transcriptome. Commun Biol 8, 915 (2025). Krizhevsky A, Sutskever I, Hinton GE. ImageNet classification with deep convolutional neural networks. Commun ACM 60, 84–90 (2017). Huang X, Hu Y, Zhang Y, Zhou Q. Two-Dimensional Ultrasound-Based Radiomics Nomogram for Diabetic Kidney Disease: A Pilot Study. Int J Gen Med 17, 1877–1885 (2024). Zhang L, et al. Preliminary study on the application of renal ultrasonography radiomics in the classification of glomerulopathy. BMC Med Imaging 21, 115 (2021). Kuo CC, et al. Automation of the kidney function prediction and classification through ultrasound-based kidney imaging using deep learning. NPJ Digit Med 2, 29 (2019). Tjoa E, Guan C. A Survey on Explainable Artificial Intelligence (XAI): Toward Medical XAI. IEEE Trans Neural Netw Learn Syst 32, 4793–4813 (2021). Rudin C. Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable Models Instead. Nat Mach Intell 1, 206–215 (2019). Bellur SS, Troyanov S, Vorobyeva O, Coppo R, Roberts ISD. Evidence from the large VALIGA cohort validates the subclassification of focal segmental glomerulosclerosis in IgA nephropathy. Kidney Int 105, 1279–1290 (2024). Rippy JR, Singh M, Aglyamov SR, Larin KV. Ultrasound Shear Wave Elastography and Transient Optical Coherence Elastography: Side-by-Side Comparison of Repeatability and Accuracy. IEEE Open Journal of Engineering in Medicine and Biology 2, 179–186 (2021). Chang TW, et al. Artificial intelligence for predicting interstitial fibrosis and tubular atrophy using diagnostic ultrasound imaging and biomarkers. BMJ Health Care Inform 32, (2025). Zhu W, et al. A clinically applicable CKD diagnostic model derived from sound touch viscosity ultrasound and LASSO regression. Front Bioeng Biotechnol 13, 1651500 (2025). Zhang X, et al. The impact of the combat method on radiomics feature compensation and analysis of scanners from different manufacturers. BMC Med Imaging 24, 137 (2024). Khaked AA, Oishi N, Roggen D, Lago P. In Shift and In Variance: Assessing the Robustness of HAR Deep Learning Models Against Variability. Sensors (Basel) 25, (2025). Zhou SK, et al. A review of deep learning in medical imaging: Imaging traits, technology trends, case studies with progress highlights, and future promises. Proc IEEE Inst Electr Electron Eng 109, 820–838 (2021). He KM, Zhang XY, Ren SQ, Sun J, Ieee. Deep Residual Learning for Image Recognition. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) ) (2016). Ioffe S, Szegedy C. Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. In: 32nd International Conference on Machine Learning ) (2015). Matias R, Jorge G-S, Victor AJ, Adrian W. Zorro: A Flexible and Differentiable Parametric Family of Activation Functions That Extends ReLU and GELU.) (2024). Qifeng Z, Barry YC, Nelson M, Andreas S. On using MLP features in LVCSR.) (2004). Qian Q, Jin R, Yi J, Zhang L, Zhu S. Efficient distance metric learning by adaptive sampling and mini-batch stochastic gradient descent (SGD). Springer US , (2015). Tables Table 1 Baseline characteristics. Characteristics Training Cohort Test Cohort External Validation Cohort IgAN (n=601) Controls (n=655) p -value IgAN (n=257) Controls (n=281) p -value IgAN (n=105) Controls (n=113) p -value Age (years, mean ± SD) 33.21 ± 11.70 34.29 ± 12.43 0.08 34.04± 12.15 32.91 ± 12.06 0.28 34.50 ± 12.20 33.80 ± 11.90 0.12 Gender (Male/Female) 327/27 4336/319 0.15 142/115 162/119 0.58 50/55 60/53 0.42 Serum Creatinine (μmol/L) 134.07 ± 18.07 80.10 ± 11.43 ＜0.001 131.43± 7.35 81.30 ± 11.02 ＜0.001 132.55 ± 17.55 80.55 ± 11.25 ＜0.001 eGFR (mL/min/1.73m²) 72.74 ± 15.43 103.31 ± 9.80 ＜0.001 70.51± 13.95 102.66 ± 11.98 ＜0.001 70.85 ± 14.55 102.05 ± 10.04 ＜0.001 Table 2 Diagnostic performance of baseline models versus IANet variants Performance metrics include accuracy, AUC, TPR, and FPR with 95% confidence intervals Model Accuracy (95% CI) AUC (95% CI) TPR (95% CI) FPR (95% CI) PPV (95% CI) NPV (95% CI) ResNet-18 Baseline 84.47% (82.35–86.32%) 0.892 (0.871–0.910) 81.3% (78.5–83.8%) 15.6% (13.2–17.9%) 82.9% (78.2–87.1%) 83.2% (78.5–87.3%) IANet (ResNet-18) 96.73% (95.81–97.49%) 0.975 (0.962–0.984) 95.2% (93.1–96.8%) 3.1% (2.3–4.0%) 96.8% (94.1–98.7%) 95.8% (92.6–97.9%) ResNet-50 Baseline 88.62% (86.85–90.18%) 0.921 (0.903–0.936) 85.7% (83.2–87.9%) 11.2% (9.5–13.1%) 87.7% (83.5–91.1%) 87.1% (82.8–90.6%) IANet (ResNet-50) 95.82% (94.73–96.69%) 0.968 (0.953–0.979) 93.8% (91.5–95.6%) 4.5% (3.6–5.4%) 94.9% (91.5–97.3%) 94.4% (90.8–96.8%) DenseNet-121 Baseline 86.35% (84.42–88.09%) 0.905 (0.884–0.923) 83.1% (80.4–85.5%) 13.5% (11.7–15.5%) 84.9% (80.4–88.7%) 84.7% (79.9–88.6%) IANet (DenseNet-121) 94.95% (93.68–95.98%) 0.961 (0.945–0.973) 92.1% (89.7–94.0%) 5.8% (4.9–6.7%) 93.7% (90.0–96.4%) 93.0% (89.1–95.8%) InceptionV3 Baseline 85.12% (83.15–86.89%) 0.898 (0.877–0.916) 82.4% (79.6–84.8%) 14.3% (12.5–16.3%) 84.1% (79.7–87.9%) 84.3% (79.8–87.9%) IANet (InceptionV3) 93.27% (91.92–94.38%) 0.952 (0.935–0.965) 89.5% (87.0–91.7%) 7.3% (6.2–8.5%) 91.6% (87.8–94.5%) 90.6% (86.7–93.6%) Table 3 Diagnostic performance of IANet in the multi-center external validation cohort Metric Category Metric Name Abbreviation Value 95% Confidence Interval (95% CI) Cohort Characteristics Total Samples - 218 - Biopsy-Confirmed IgAN Cases - 105 - Non-IgAN Controls (including healthy/other CKD) - 113 - Core Diagnostic Efficacy Area Under ROC Curve AUC 0.9275 0.8912–0.9548 Overall Accuracy Acc 0.8624 0.8115–0.9037 Confusion Matrix True Positive (IgAN correctly classified) TP 101 - True Negative (Non-IgAN correctly classified) TN 87 - False Positive (Non-IgAN misclassified as IgAN) FP 26 - False Negative (IgAN misclassified as Non-IgAN) FN 4 - Clinical Risk Metrics Sensitivity (TPR, avoiding missed diagnosis) Sen 0.9619 0.9157–0.9864 Specificity (1-FPR, avoiding misdiagnosis) Spe 0.7700 0.6853–0.8398 Diagnostic Total Diagnostic Time (for 218 samples) - 8.85 s - Table 4 Diagnostic Performance Comparison Evaluator Accuracy (95% CI) PPV (95% CI) NPV (95% CI) diagnostic Time (s) IANet 98.0% (95.3-100.0%) 100.0% (91.2-100.0%) 96.7% (88.7-99.6%) 4 Radiologist X. G 82.0% (74.4-89.6%) 80.0% (64.3-90.9%) 83.3% (71.5-91.7%) 2040 Radiologist J. S 74.0% (65.5-82.5%) 94.4% (72.7-99.9%) 69.5% (58.7-78.9%) 1680 Radiologist F. H 71.0% (62.2-79.8%) 84.2% (60.4-96.6%) 67.9% (57.1-77.5%) 2160 Table 5 Correlation Analysis of IANet’s Attention Patterns with Histopathological Features Pathological/Clinical Feature Correlation Measure 95% CI p -value Sample Size (n) Analysis Method Histopathological Correlation Mesangial proliferation (M1) - - - 18(5.0%) excluded, Segmental sclerosis (S1) OR = 4.28 2.95–6.21 <0.001 198(54.7%) Multivariate logistic regression Endothelial proliferation (E1) OR = 2.15 1.32–3.51 0.002 97(26.8%) Multivariate logistic regression Crescent formation (C1&C2) OR = 1.83 0.97–3.45 0.063 114(31.5%) Multivariate logistic regression Interstitial fibrosis/tubular atrophy (T1&T2) OR = 3.20 2.25 – 4.55 <0.001 129(35.6%) Multivariate logistic regression Imaging-Pathology Quantitative Correlation Cortical echogenicity increase r = 0.69 0.59–0.77 <0.001 87 Pearson correlation Notes: The independent variable for all logistic regression models was renal parenchyma gray level co-occurrence matrix-angular second moment (GLCM-ASM) (lower GLCM-ASM values indicate higher parenchymal structural disorganization and corresponding model attention heterogeneity). All logistic regression models were adjusted for age and gender. Abbreviations: OR = Odds Ratio; CI = Confidence Interval. Additional Declarations There is NO Competing Interest. Supplementary Files Supplementarytables.docx Supplementary table supplementarfigure1.png Supplementary figure 1 Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8425698","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":582648005,"identity":"cf891738-3a9c-41f1-babd-8df0d3d19677","order_by":0,"name":"Dong Wang","email":"data:image/png;base64,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","orcid":"","institution":"University of Electronic Science and Technology of China","correspondingAuthor":true,"prefix":"","firstName":"Dong","middleName":"","lastName":"Wang","suffix":""},{"id":582648006,"identity":"3425c4a8-6c20-485b-86da-83a0cc1d4d58","order_by":1,"name":"Yanyan Wu","email":"","orcid":"","institution":"University of Electronic Science and Technology of China","correspondingAuthor":false,"prefix":"","firstName":"Yanyan","middleName":"","lastName":"Wu","suffix":""},{"id":582648007,"identity":"97ccc99a-16f7-422b-bfef-f7a9b9268d18","order_by":2,"name":"JinQi Gong","email":"","orcid":"","institution":"University of Electronic Science and Technology of China","correspondingAuthor":false,"prefix":"","firstName":"JinQi","middleName":"","lastName":"Gong","suffix":""},{"id":582648008,"identity":"a4f57757-027a-4bda-8ba5-cbfe363cec79","order_by":3,"name":"ZiHan Tang","email":"","orcid":"","institution":"University of Electronic Science and Technology of China","correspondingAuthor":false,"prefix":"","firstName":"ZiHan","middleName":"","lastName":"Tang","suffix":""},{"id":582648009,"identity":"4c2923e7-07ae-4b23-8a18-4a7461b985b6","order_by":4,"name":"MingRui Fu","email":"","orcid":"","institution":"University of Electronic Science and Technology of China","correspondingAuthor":false,"prefix":"","firstName":"MingRui","middleName":"","lastName":"Fu","suffix":""},{"id":582648010,"identity":"8d6587f2-f9c5-41d5-94ac-f1210e71802d","order_by":5,"name":"Yunhao Luo","email":"","orcid":"","institution":"Department of Ultrasound, Sichuan Integrative Medicine Hospital","correspondingAuthor":false,"prefix":"","firstName":"Yunhao","middleName":"","lastName":"Luo","suffix":""},{"id":582648011,"identity":"71b9ca4d-d724-40fe-bbe5-8060a8e660a8","order_by":6,"name":"ChengYuan Ye","email":"","orcid":"","institution":"University of Electronic Science and Technology of China","correspondingAuthor":false,"prefix":"","firstName":"ChengYuan","middleName":"","lastName":"Ye","suffix":""},{"id":582648012,"identity":"8c9b23f6-40f0-4ce6-8824-96ca63bb7dc5","order_by":7,"name":"XinYi Hu","email":"","orcid":"","institution":"University of Electronic Science and Technology of China","correspondingAuthor":false,"prefix":"","firstName":"XinYi","middleName":"","lastName":"Hu","suffix":""},{"id":582648013,"identity":"c5d5f5a1-0846-48ac-b0fc-f999634eaf73","order_by":8,"name":"Xueying Chen","email":"","orcid":"","institution":"University of Electronic Science and Technology of China","correspondingAuthor":false,"prefix":"","firstName":"Xueying","middleName":"","lastName":"Chen","suffix":""},{"id":582648014,"identity":"68b6d206-8835-444c-8a1c-9fed6f28b148","order_by":9,"name":"Yong Jin","email":"","orcid":"","institution":"The First Affiliated Hospital of Ningbo University","correspondingAuthor":false,"prefix":"","firstName":"Yong","middleName":"","lastName":"Jin","suffix":""},{"id":582648015,"identity":"81b2d8c7-49b4-427a-89a1-a8d1fdb7ce0a","order_by":10,"name":"Ping Zhang","email":"","orcid":"","institution":"Nephrology Department, ,Sichuan Provincial People's Hospital, University of Electronic Science and Technology of China","correspondingAuthor":false,"prefix":"","firstName":"Ping","middleName":"","lastName":"Zhang","suffix":""}],"badges":[],"createdAt":"2025-12-22 14:12:14","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8425698/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8425698/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":101485486,"identity":"466abb72-5cca-441f-9290-a4679648507d","added_by":"auto","created_at":"2026-01-30 09:00:43","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":352122,"visible":true,"origin":"","legend":"\u003cp\u003ePatient selection and dataset partitioning flowchart. A total of 2,012 participants from four Chinese institutions were enrolled: 1,794 from Sichuan Provincial People’s Hospital (primary centre) split into training (n=1,256) and internal test (n=538) sets; 218 from three external centres (validation cohort). Inclusion/exclusion criteria are detailed in the Methods section.\u003c/p\u003e","description":"","filename":"fig1.png","url":"https://assets-eu.researchsquare.com/files/rs-8425698/v1/d3d4270ab4b1946e99909221.png"},{"id":101485492,"identity":"dbf7660b-3c18-4492-8b64-620460f7b3f8","added_by":"auto","created_at":"2026-01-30 09:00:43","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":3037820,"visible":true,"origin":"","legend":"\u003cp\u003eIANet development and evaluation flowchart. (a) Data curation: 3,588 bilateral kidney ultrasound images (1,794 participants) linked to biopsy reports; preprocessing (resizing, augmentation) and quality control (image score ≥3). (b) Model training (ResNet-18 backbone, 300 epochs) and multi-stage evaluation: internal validation (n=538), external validation (n=218), radiologist benchmarking (n=100), and pathological/biomarker correlation (n=234).\u003c/p\u003e","description":"","filename":"fig2.png","url":"https://assets-eu.researchsquare.com/files/rs-8425698/v1/6c3140eafb1c4dcdc540b29d.png"},{"id":101751943,"identity":"bcbc5167-dc7d-4f8e-8f66-57cb451783ef","added_by":"auto","created_at":"2026-02-03 10:24:30","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":3084352,"visible":true,"origin":"","legend":"\u003cp\u003eRepresentative ultrasound images and IANet attention heatmaps. (a) Healthy control (33-year-old female): Normal renal parenchyma, clear corticomedullary junction; IANet attention focused on normal cortex (green heatmap). (b) IgAN patient (19-year-old male, M1E1S1T0C1): Increased cortical echogenicity (red arrow) vs. liver parenchyma; IANet attention concentrated on cortical lesions (red heatmap). (c) DKD patient (56-year-old male): Mild parenchymal heterogeneity (yellow arrow); IANet attention on corticomedullary junction (orange heatmap). Scale bar: 1 cm.\u003c/p\u003e","description":"","filename":"fig3.png","url":"https://assets-eu.researchsquare.com/files/rs-8425698/v1/84873d7d7bf61dd74cb47c2b.png"},{"id":101485489,"identity":"9ab473cb-9c43-43c6-a4d9-55d1006d8452","added_by":"auto","created_at":"2026-01-30 09:00:43","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":729058,"visible":true,"origin":"","legend":"\u003cp\u003eIANet architecture diagram (labeled signal flow). Modules: (1) Shared Feature Extraction (ResNet-18 backbone); (2) Individual Feature Branches (channel attention for left/right kidneys); (3) Cross-Interaction (multi-head attention for bidirectional feature exchange); (4) Enhanced Difference (asymmetry amplification); (5) Multi-Dimensional Fusion + Classification. Blue arrows indicate signal flow between modules; numbers in parentheses denote feature dimensions.\u003c/p\u003e","description":"","filename":"fig4.png","url":"https://assets-eu.researchsquare.com/files/rs-8425698/v1/dd3c1836ee069a8a974a9611.png"},{"id":101485491,"identity":"0e8e5563-d480-44cd-a4a5-a839569c6eb0","added_by":"auto","created_at":"2026-01-30 09:00:43","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":4034994,"visible":true,"origin":"","legend":"\u003cp\u003eGrad-CAM reveals spectrum of IANet attention patterns, correlating with IgAN pathology. From red (high attention, \u0026gt;0.7) to blue (low attention, \u0026lt;0.3). (a) In a healthy control (34F), attention is uniformly distributed. (b) A confirmed IgAN case (19M, M1E1S1T0C1) shows markedly heterogeneous attention. (c) A subtle IgAN case (36F, M0E1S0T0C1), missed by radiologists, presented low attention heterogeneity, which IANet nonetheless detected. (d) An ultra-early IgAN case (45M, M0E0S0T0C0, IFTA\u0026lt;5%) with homogeneous attention was missed by the model. (e) Representative S1T2-stage IgAN case (33M, M0E0S1T2C0): (e1) PAS-stained renal biopsy (segmental sclerosis, severe interstitial fibrosis); (e2) Renal ultrasound; (e3) Parenchymal ROI; (e4) Patchy heterogeneous attention heatmap; (e5–e6) Parenchymal ROI masks; (e7) Directional GLCM-ASM distribution (avg=0.000785) matching attention heterogeneity and pathology.\u003c/p\u003e","description":"","filename":"fig5.png","url":"https://assets-eu.researchsquare.com/files/rs-8425698/v1/fbf199fb7df02bd31d715ecd.png"},{"id":105566359,"identity":"4cd2d111-8e3a-43e1-8664-e25098268d1e","added_by":"auto","created_at":"2026-03-27 12:56:16","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":20651958,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8425698/v1/95fafd33-1f8e-4dae-9566-67e7326bd7e6.pdf"},{"id":101485487,"identity":"9af9761b-5a62-4f94-96b4-1105530f50a9","added_by":"auto","created_at":"2026-01-30 09:00:43","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":36666,"visible":true,"origin":"","legend":"\u003cp\u003eSupplementary table\u003c/p\u003e","description":"","filename":"Supplementarytables.docx","url":"https://assets-eu.researchsquare.com/files/rs-8425698/v1/0a16e098ddd3bb89b14b86ac.docx"},{"id":101485488,"identity":"fcd2b3b0-5df0-45e3-836b-0ab32c595890","added_by":"auto","created_at":"2026-01-30 09:00:43","extension":"png","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":700187,"visible":true,"origin":"","legend":"\u003cp\u003eSupplementary figure 1\u003c/p\u003e","description":"","filename":"supplementarfigure1.png","url":"https://assets-eu.researchsquare.com/files/rs-8425698/v1/8ffbe825258b9e9e8fc43d53.png"}],"financialInterests":"There is \u003cb\u003eNO\u003c/b\u003e Competing Interest.","formattedTitle":"A Deep Learning Framework Based on Ultrasomics for IgA Nephropathy Detection: A Multi-Centre, Multi-Scanner Validation Study","fulltext":[{"header":"Introduction","content":"\u003cp\u003eImmunoglobulin A nephropathy (IgAN), characterized by dominant glomerular IgA deposition, is the leading cause of primary glomerulonephritis globally and accounts for 45% of such diagnoses in China\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. Despite therapeutic advances (e.g., immunosuppressants and renin-angiotensin system inhibitors), \u0026gt;\u0026thinsp;30% of patients progress to end-stage kidney disease (ESKD) within 10\u0026ndash;20 years of diagnosis\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e, imposing substantial clinical and economic burdens (e.g., \u0026gt;\u003cspan\u003e$\u003c/span\u003e2.4\u0026nbsp;billion annual ESKD management costs in the US alone\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eThe gold standard for IgAN diagnosis\u0026mdash;renal biopsy with the Oxford MEST-C score\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e, is limited by three key barriers: (1) invasiveness, with a 2\u0026ndash;5% risk of major complications (e.g., perirenal hematoma)\u003csup\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e; (2) variable regional biopsy thresholds\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e, leading to missed early intervention opportunities; (3) lack of integration with imaging biomarkers, as the MEST-C score relies solely on histology and clinical metrics. Renal ultrasound, the preferred non-invasive modality for chronic kidney disease (CKD), detects parenchymal echogenicity increases linked to renal fibrosis\u0026mdash;a stronger predictor of IgAN outcomes than traditional biomarkers (e.g., proteinuria)\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e. However, this clinician-dependent approach introduces substantial inter-observer variability, compromising diagnostic reliability\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e (inter-observer agreement: Fleiss\u0026rsquo; κ\u0026thinsp;=\u0026thinsp;0.366 in our study) and insensitivity to early lesions \u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eArtificial intelligence (AI) has emerged to enhance ultrasound diagnostics\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e, but three critical barriers hinder clinical translation: (1) poor generalizability across heterogeneous scanners and acquisition protocols\u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e, as most models are trained on single-centre, single-device datasets; (2) scarce histopathologically annotated data\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e, introducing population bias (e.g., overfitting to Asian cohorts)\u003csup\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e; (3) inability to detect subtle IgAN-specific microstructural changes amid anatomical noise from age-related or comorbidity-related alterations\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e. Prior AI models for IgAN either rely on manual region-of-interest (ROI) delineation\u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e, which is a time-consuming step incompatible with clinical workflows, or focus on single-kidney feature extraction\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e without leveraging the physiological symmetry of bilateral kidneys, and this symmetry serves as a key insight for amplifying pathological differences.\u003c/p\u003e \u003cp\u003eTo address these gaps, we developed \u003cb\u003eIANet\u003c/b\u003e, a deep learning framework that: (1) uses BRSA to enhance detection of early, asymmetric lesions; (2) enables fully automated kidney localization (eliminating manual ROI bias); (3) ensures interpretability via Grad-CAM (validated against histopathology and biomarkers, Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e5\u003c/span\u003e); (4) maintains robustness across multi-centre, multi-scanner datasets. Our results demonstrate IANet is a potential non-invasive tool to provide imaging support for IgAN disease course\u003c/p\u003e "},{"header":"Results","content":"\u003cp\u003e\u003cstrong\u003e1. Study Population and Baseline Characteristics\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFigure 1 outlines the study flowchart A total of 2,012 participants (four institutions in China) were enrolled: 963 biopsy-confirmed IgAN patients (Oxford MEST-C stages: M0/M1=245/511, E0/E1=623/131, S0/S1=589/167, T0/T1/T2=328/287/131, C0/C1-C2=652/104) and 1,049 healthy controls (normal renal function, no hematuria/hypertension). The cohort included 1,077 males (53.53%) and 935 females (46.47%), with a mean age of 34.07 \u0026plusmn; 12.05 years (range: 18\u0026ndash;78 years). \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eData from the primary centre (Sichuan Provincial People\u0026rsquo;s Hospital, Chengdu, n=1,794) were split into training (n=1,256: 601 IgAN, 655 controls) and internal test sets (n=538: 257 IgAN, 281 controls) at a 7:3 ratio using stratified random sampling (preserving IgAN stage distribution). As shown in Figure 2, the development process of IANet encompasses data curation, model training (Figure 2a), and multi-stage evaluation (Figure 2b). The external validation cohort (n=218: 105 IgAN, 113 controls) included participants from three independent centres: The First Affiliated Hospital of Chongqing Medical University (Chongqing), Sichuan Integrative Medicine Hospital (Chengdu), and The First Affiliated Hospital of Ningbo University (Ningbo). \u0026nbsp;All data were analyzed using ultrasound images of the maximal longitudinal sections of bilateral kidneys (Figure 3).\u003c/p\u003e\n\u003cp\u003eBaseline demographic and clinical characteristics (Table 1)\u0026nbsp;showed no significant differences in age or sex between the IgA nephropathy (IgAN) group and the control group across the training, internal test, and external validation cohorts. However, significant differences were observed in key clinical biomarkers (all p \u0026lt; 0.001). In all three cohorts, serum creatinine levels were higher in the IgAN group compared to the control group, while the eGFR was lower in the IgAN group. These findings are consistent with the known pathophysiological characteristics of IgA nephropathy. To ensure full transparency regarding the imaging data sources, the sample distribution across all participating centres and ultrasound manufacturers is detailed in Supplementary Table 1. This distribution reflects real-world clinical usage patterns. Briefly, the four major manufacturers (Mindray, Philips, Samsung, and GE) collectively contributed 84.2% of the total dataset (1,695/2,012 samples), with Mindray being the single most prevalent brand (28.2%).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2. Model Performance: Baseline vs. IANet\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIANet outperformed four baseline architectures (ResNet-18, ResNet-50, DenseNet-121, InceptionV3) across all metrics (all p\u0026lt;0.001; Table 2). In internal validation: \u0026nbsp;\u003cstrong\u003eResNet-18-based IANet\u003c/strong\u003e (prioritized for clinical translation due to balance of performance and efficiency) achieved 96.73% accuracy (95% CI: 95.81\u0026ndash;97.49%) and 0.975 AUC (95% CI: 0.962\u0026ndash;0.984)\u0026mdash;12.26% higher accuracy and 0.083 higher AUC than baseline ResNet-18 (84.47% accuracy, 0.892 AUC). \u0026nbsp;False positive rate (FPR) was reduced by 76.80% vs. baseline ResNet-18 (3.1% [95% CI: 2.3\u0026ndash;4.0%] vs. 15.6% [95% CI: 13.2\u0026ndash;17.9%]), minimizing unnecessary follow-ups. \u0026nbsp;Positive predictive value (PPV) = 96.8% (95% CI: 94.1\u0026ndash;98.7%) and negative predictive value (NPV) = 95.8% (95% CI: 92.6\u0026ndash;97.9%), confirming reliability in both IgAN identification (high PPV) and rule-out (high NPV). \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eResNet-50/DenseNet-121/InceptionV3-based IANet variants also showed significant improvements over their baselines (e.g., ResNet-50: 88.62% \u0026rarr; 95.82% accuracy, p\u0026lt;0.001), but ResNet-18-based IANet had the lowest computational cost (inference time: 0.04 s per case vs. 0.12 s for ResNet-50-based IANet) and highest generalizability (external validation AUC: 0.9275 vs. 0.901 for ResNet-50-based IANet; \u0026nbsp;Supplementary Table 2). \u0026nbsp;\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e1. Innovation Over Existing Research\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIANet addresses three unmet needs in IgAN diagnostics, distinguishing it from prior AI models: \u0026nbsp;\u003cstrong\u003eBRSA for asymmetry amplification\u003c/strong\u003e: Unlike single-kidney models\u003csup\u003e15, 22\u003c/sup\u003e, BRSA amplifies pathological asymmetries (e.g., subtle cortical echogenicity differences between left/right kidneys) while suppressing symmetric anatomical noise (e.g., age-related parenchymal changes). Ablation studies (Table 2) confirm BRSA drives a 12.26% accuracy gain in ResNet-18 variants\u0026mdash;critical for detecting early T0-stage lesions. \u003cstrong\u003eFully automated localization\u003c/strong\u003e: Eliminates manual ROI delineation \u003csup\u003e23\u003c/sup\u003e, which is a bottleneck in prior studies\u003csup\u003e17\u003c/sup\u003e. Grad-CAM and Table 5 confirm IANet accurately identifies kidneys in complex abdominal scans (e.g., amid liver/spleen) and focuses on pathologically relevant regions. This automation reduces inter-operator variability and enables real-time clinical integration. \u003cstrong\u003eMulti-modal interpretability\u003c/strong\u003e: Unlike \u0026ldquo;black-box\u0026rdquo; AI\u003csup\u003e24, 25\u003c/sup\u003e, IANet provides interpretability by linking its attention heterogeneity directly to underlying pathology. This heterogeneity, quantified as GLCM-ASM was significantly associated with chronic fibrotic lesions (e.g., S1, T1\u0026amp;T2)\u003csup\u003e26\u003c/sup\u003e, demonstrating that the model detects IgAN by capturing parenchymal structural disarray, thereby building crucial clinical trust. Compared to published IgAN AI models (Supplementary Table 6), IANet achieves higher external validation AUC (0.9275 vs. 0.82\u0026ndash;0.89) and faster processing (0.04 s vs. 1\u0026ndash;5 s per case), with the unique advantage of bilateral analysis.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3. IANet Performance in External Validation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIn external validation (Table 3), IANet maintained robust performance across heterogeneous scanners (6 manufacturers: GE, Philips, Samsung, Mindray, Toshiba, Canon) and centres: \u0026nbsp;Accuracy = 86.24% (95% CI: 81.15\u0026ndash;90.37%), AUC = 0.9275 (95% CI: 0.8912\u0026ndash;0.9548%); Sensitivity = 96.19% (95% CI: 91.57\u0026ndash;98.64%) (minimizing missed diagnoses), specificity = 77.00% (95% CI: 68.53\u0026ndash;83.98%); \u0026nbsp;Confusion matrix: True Positives (TP)=101, True Negatives (TN)=87, False Positives (FP)=26, False Negatives (FN)=4; \u0026nbsp; Processing time for 218 samples = 8.85 s (0.04 s per case), with a false negative rate of 1.83% (4/105 IgAN cases misclassified, all T0 stage with IFTA \u0026lt;5%). \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCross-device performance\u003c/strong\u003e (Supplementary Table 2) confirmed no significant differences in accuracy across major manufacturers (F=1.23, p=0.30). The lower overall specificity (77.00%) vs. individual device specificity (92.9\u0026ndash;94.1%) was attributed to 26 false positives from two low-volume manufacturers, which were included to reflect real-world device diversity. \u0026nbsp;Prior multi-center studies have reported significant performance drops (5-12% accuracy reduction) when models are tested across heterogeneous scanner types\u003csup\u003e17\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4. Ablation Study: Impact of Bilateral Interaction\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAblation experiments compared full IANet (with BRSA modules: cross-interaction, enhanced difference, multi-dimensional fusion) to baseline models (concatenated bilateral inputs without BRSA). The full model showed significant improvements (Table 2): \u0026nbsp;ResNet-18-based IANet: Accuracy 84.47% \u0026rarr; 96.73% (+12.26%, 95% CI: 10.12\u0026ndash;14.18%), AUC 0.892 \u0026rarr; 0.975 (+0.083), FPR 15.6% \u0026rarr; 3.1% (80% reduction); ResNet-50-based IANet: FPR 11.2% \u0026rarr; 4.5% (59.8% reduction), sensitivity 85.7% \u0026rarr; 93.8% (+8.1%); \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThese results confirm that \u003cstrong\u003ecross-interaction and enhanced difference modules\u003c/strong\u003e (core of BRSA) drive performance gains by amplifying pathological asymmetries (e.g., subtle cortical echogenicity differences between left/right kidneys) while suppressing symmetric anatomical noise.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e5. IANet vs. Radiologists\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIn a blinded assessment of 100 cases (42 IgAN, 58 controls; image quality score \u0026ge;4/5, excluding motion artifacts), Radiologists were provided only with the de-identified ultrasound images and were blinded to the patients\u0026apos; clinical history, laboratory results (e.g., eGFR), and biopsy findings, to ensure a direct comparison of visual feature extraction capabilities against the IANet model. IANet outperformed three board-certified radiologists (10\u0026ndash;15 years of nephrology ultrasound experience; Table 4): \u0026nbsp; Accuracy: IANet 98.0% (95% CI: 95.3\u0026ndash;100.0%) vs. Radiologist X.G (15-yr exp) 82.0% (95% CI: 74.4\u0026ndash;89.6%), Radiologist J.S (10-yr exp) 74.0% (95% CI: 65.5\u0026ndash;82.5%), Radiologist F.H (12-yr exp) 71.0% (95% CI: 62.2\u0026ndash;79.8%) (all p\u0026lt;0.001, Cochran\u0026rsquo;s Q test + post-hoc McNemar with Bonferroni adjustment, \u0026alpha;=0.0083,\u0026nbsp;supplementary table 3); \u0026nbsp; Predictive values: IANet PPV=100% (no false positives), NPV=96.7% vs. radiologists PPV=68.9\u0026ndash;79.4%, NPV=67.9\u0026ndash;83.3%; \u0026nbsp; Processing time: IANet=4 s vs. radiologists=28\u0026ndash;36 min (\u0026gt;1,000-fold faster, critical for high-volume clinical settings). \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eConsistent with the accuracy results, the receiver operating characteristic (ROC) curve analysis further validated IANet\u0026apos;s superior discriminative capability. In the internal validation cohort, IANet (ResNet-18) achieved an AUC of 0.975 (95% CI: 0.962\u0026ndash;0.984), which was significantly higher than that of the ResNet-18 baseline (0.892, 95% CI: 0.871\u0026ndash;0.910) and ResNet-50 baseline (0.921, 95% CI: 0.903\u0026ndash;0.936), as well as the reported AUC of conventional ultrasound (0.820, 95% CI: 0.79\u0026ndash;0.85) (Supplementary Figure 1A). In the blinded 100-case assessment, IANet exhibited an outstanding AUC of 0.992 (95% CI: 0.978\u0026ndash;1.000), substantially outperforming all three radiologists (Radiologist X.G: 0.864, 95% CI: 0.792\u0026ndash;0.936; Radiologist J.S: 0.782, 95% CI: 0.700\u0026ndash;0.864; Radiologist F.H: 0.751, 95% CI: 0.665\u0026ndash;0.837) (Supplementary Figure 1B).\u003c/p\u003e\n\u003cp\u003eInter-rater agreement among radiologists was substantial (Fleiss\u0026rsquo; \u0026kappa;=0.366, 95% CI: 0.253\u0026ndash;0.480; p\u0026lt;0.001; supplementary table 4), highlighting subjective interpretation bias\u0026mdash;an issue IANet eliminates via automated, objective analysis. IANet\u0026rsquo;s superior performance is supported by its specialized architecture designed for bilateral renal ultrasound analysis (Figure 4), which integrates bilateral symmetry-aware modules to capture pathological asymmetries critical for IgAN detection\u003c/p\u003e\n\u003cp\u003eWe applied gradient-weighted class activation mapping (Grad-CAM) to visualize spatial attention patterns in IANet\u0026rsquo;s decision-making process. As shown in Figure 5, heatmaps (red/orange = high attention; green/blue = low attention) reveal the model\u0026rsquo;s diagnostic rationale through region-specific feature activation. Consistent anatomical focus in the Grad-CAM heatmap confirms that IANet can automatically locate kidneys in ultrasound images containing complex anatomical structures such as liver, spleen, and subcutaneous fat, achieving precise spatial alignment with ultrasound-defined renal borders (Fig. 5). Meanwhile, IANet exhibits attention to adjacent organs (liver/spleen), indicating that this may be consistent with radiologists\u0026rsquo;\u0026nbsp;diagnostic thinking of contrasting the kidneys with surrounding structures.\u003c/p\u003e\n\u003cp\u003eAll diagnostic errors occurred exclusively in early Oxford T0-stage IgAN (interstitial fibrosis and tubular atrophy [IFTA] \u0026lt;25%): \u0026nbsp;IANet missed 2 cases (2% error rate), both with minimal fibrosis (IFTA \u0026lt;5%; one with predominant podocyte injury, no detectable echogenicity changes); \u0026nbsp; Radiologists missed 10\u0026ndash;26 cases (17.2\u0026ndash;44.8% error rates), primarily T0-stage cases with IFTA 5\u0026ndash;15%; \u0026nbsp;Eight cases were missed by all evaluators (including IANet\u0026rsquo;s 2 cases), all with IFTA \u0026lt;15%\u0026mdash;indicating a shared diagnostic blind spot for ultra-early lesions with non-specific sonographic features.\u0026nbsp;This observation that misclassifications were confined solely to early-stage disease prompted us to systematically evaluate the model\u0026rsquo;s performance across the full spectrum of Oxford MEST-C pathological phenotypes in both internal and external validation cohorts.\u003c/p\u003e\n\u003cp\u003eWith only 18 M1 cases (5.0% of the combined internal and external validation cohort, n=362 [257+105])) in the M subgroup; insufficient sample size limits statistical power for reliable subgroup analysis, thus, the M subgroup is excluded, with only E, S, T, C included for subsequent analyses. The results demonstrated that IANet maintained high accuracy for advanced IgAN stages but showed reduced performance for ultra-early disease (Supplementary Table 5): \u0026nbsp;High-performance stages: S1 (97.0% [95% CI: 93.6\u0026ndash;98.8%]), T1-T2 (94.8\u0026ndash;96.2%). Low-performance stage:\u0026nbsp;T0 (IFTA \u0026lt;25%: 83.6% [95% CI: 79.2\u0026ndash;87.3%]), primarily due to minimal fibrosis (\u0026lt;5% IFTA) in 12 misclassified cases.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e6. Correlation of IANet Attention with Histopathology\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAnalysis of Grad-CAM attention maps revealed that IgA nephropathy (IgAN) cases with chronic lesions presented significantly higher attention distribution heterogeneity in the renal parenchyma. Based on this observation, we hypothesized that the degree of IANet\u0026rsquo;s attention heterogeneity correlates with lesion pathological subtypes, and verified this association using texture uniformity analysis based on gray level co-occurrence matrix-angular second moment (GLCM-ASM) (Table 5).\u003c/p\u003e\n\u003cp\u003eParenchymal attention heterogeneity showed strong, significant associations with chronic fibrosing pathologies. It served as a powerful independent predictor for segmental sclerosis (S1) (OR=4.28, 95% CI: 2.95\u0026ndash;6.21, p\u0026lt;0.001, n=198, 54.7% of validation cohort) and interstitial fibrosis/tubular atrophy (T1\u0026amp;T2) (OR=3.20, 95% CI: 2.25\u0026ndash;4.55, p\u0026lt;0.001, n=129, 35.6% of validation cohort). In contrast, its associations with acute lesions were markedly weaker: a significant but modest link was found with endothelial proliferation (E1) (OR=2.15, 95% CI: 1.32\u0026ndash;3.51, p=0.002, n=97, 26.8% of validation cohort), while the association with crescent formation (C1\u0026amp;C2) showed a positive trend without reaching statistical significance (OR=1.83, 95% CI: 0.97\u0026ndash;3.45, p=0.063, n=114, 31.5% of validation cohort). Mesangial proliferation (M1) was excluded from analysis due to insufficient sample size (n=18, 5.0%).\u003c/p\u003e\n\u003cp\u003eAs shown in Figure 5, renal parenchyma in healthy controls (Figure 5a) exhibited homogeneous attention distribution. Conversely, correctly classified IgAN cases (Figure 5b) presented high attention heterogeneity, with focus concentrated on regions of abnormal cortical echogenicity. A challenging T0-stage case (IFTA 10\u0026ndash;15%) missed by all radiologists but identified by IANet (Figure 5c) demonstrated the model\u0026rsquo;s capacity to detect subtle, low-heterogeneity pathological cues. Meanwhile, the model\u0026rsquo;s misdiagnosis of an ultra-early T0-stage case (Figure 5d) was characterized by homogeneous attention, consistent with the absence of sonographically discernible fibrosis. A representative S1T2 case (Figure 5e) displayed patchy, heterogeneous attention patterns that matched the distribution of chronic pathological lesions.\u003c/p\u003e\n\u003cp\u003eAdditionally, cortical echogenicity elevation showed a strong linear correlation with fibrosis area ratio (r=0.69, 95% CI: 0.59\u0026ndash;0.77, p\u0026lt;0.001, n=87)\u003csup\u003e13\u003c/sup\u003e. These results confirm that IANet\u0026rsquo;s diagnostic capability originates from its sensitivity to parenchymal structural disarray, enabling effective capture of chronic sclerotic and fibrotic damage while showing limited sensitivity to acute inflammatory changes.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e7. Specificity Assessment Against Diabetic Kidney Disease\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo evaluate differential diagnostic capability, IANet was tested on a biopsy-proen DKD cohort (n=192, excluded from model training; mean age 56.2\u0026plusmn;9.8 years, 58% male, eGFR 68.3\u0026plusmn;15.2 mL/min/1.73m\u0026sup2;): \u0026nbsp; IANet correctly classified 123 DKD cases as non-IgAN (specificity=64.1%), 69 cases (35.9%) were misclassified as IgAN\u0026mdash;attributed to shared sonographic features (e.g., cortical echogenicity increase due to fibrosis) between IgAN and DKD. Subgroup analysis showed misclassified DKD cases had higher fibrosis scores (IFTA \u0026ge;25%: 78% vs. 32% in correctly classified cases, p\u0026lt;0.001)\u0026mdash;highlighting the need for multimodal refinement (e.g., integrating HbA1c or urinary biomarkers). \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe specificity of IANet for differentiating IgAN from DKD (64.1%) is comparable to or higher than previously reported ultrasound AI models\u003csup\u003e15, 27\u003c/sup\u003ebut remains suboptimal. The 35.9% false positive rate is mainly due to shared sonographic features (e.g., cortical fibrosis-induced echogenicity increase) between the two diseases. Future refinement should integrate multimodal data (e.g., HbA1c, urinary biomarkers) to enhance differential diagnostic capability, as demonstrated by recent studies showing that combining ultrasound and clinical parameters can improve DKD-IgAN discrimination to 72.8% specificity \u003csup\u003e28\u003c/sup\u003e.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003e\u003cstrong\u003e2. Clinical Translation Potential\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIANet\u0026rsquo;s design aligns with unmet clinical needs across diverse healthcare settings: \u0026nbsp;\u003cstrong\u003ePre-biopsy screening\u003c/strong\u003e: High NPV (96.7%) enables reliable IgAN rule-out, reducing unnecessary biopsies by 32% in simulated cohorts (consistent with internal validation data). For example, in patients with isolated microscopic hematuria (a common IgAN presentation), IANet could identify low-risk cases (probability \u0026lt;20%) that avoid biopsy\u0026mdash;reducing complication risks and healthcare costs. \u003cstrong\u003eMulti-centre standardization\u003c/strong\u003e: Cross-device performance (Supplementary Table 2) confirms IANet works across mainstream ultrasound manufacturers (GE, Philips, Samsung, Mindray) with no significant accuracy differences (p=0.30). This standardization addresses the problem of variable ultrasound interpretation across centres\u0026mdash;critical for multi-centre trials and global adoption. \u0026nbsp;\u003cstrong\u003eEarly intervention enabler\u003c/strong\u003e: The model\u0026rsquo;s high sensitivity (96.19%) in external validation ensures minimal missed diagnoses, while its ability to detect M1/S1 stages (accuracy \u0026gt;96%) enables early initiation of renoprotective therapy, which is crucial to reducing ESKD progression. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3. Limitations and Future Directions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eUltra-early disease (T0 stage):\u003c/strong\u003e The reduced accuracy in T0-stage IgAN (83.6%) is attributed to the inherent limitation of BRSA: ultra-early lesions (IFTA \u0026lt;25%) typically exhibit bilateral symmetry (e.g., uniform mild cortical echogenicity), making it difficult for the cross-interaction and enhanced difference modules to amplify pathological disparities. This is consistent with our histopathological findings that T0-stage cases have minimal structural disorganization (higher GLCM-ASM values, mean=0.0012 vs. 0.00078 in T1/T2 stages), resulting in insufficient discriminative features. Future work will integrate quantitative ultrasound elastography\u003csup\u003e13\u003c/sup\u003e (to measure tissue stiffness) and integrate relative clinical parameters to improve ultra-early detection\u003csup\u003e29\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDKD specificity\u003c/strong\u003e: 35.9% false positives highlight overlapping sonographic features between IgAN and DKD (e.g., fibrosis). Multimodal fusion with genomic data\u003csup\u003e21\u003c/sup\u003e or clinical parameters\u003csup\u003e30\u003c/sup\u003e will enhance differential diagnosis. \u0026nbsp;We are currently expanding our research to develop a predictive model that incorporates multimodal ultrasound\u003csup\u003e31\u003c/sup\u003e and key clinical parameters to improve diagnostic capability.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePathological spectrum bias:\u003c/strong\u003e The retrospective design of our study inherently limited the number of available M1-grade cases (n=18, 5.0%), constraining the analysis of its correlation with IANet\u0026apos;s attention. Consequently, our pathological validation primarily underscores the model\u0026apos;s strength in identifying chronic, fibrotic changes (S1, T1\u0026amp;T2), while the association with active inflammatory lesions requires future prospective studies with targeted enrollment.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDevice-related variability\u003c/strong\u003e: Lower specificity in Toshiba/Canon devices (69.2\u0026ndash;71.4%, Supplementary Table 2) is attributed to differences in image resolution (512\u0026times;512 vs. 1024\u0026times;1024 for GE/Philips) and acquisition protocols (e.g., gain settings)\u003csup\u003e32, 33\u003c/sup\u003e. We are developing a device-specific calibration module (based on transfer learning) to standardize image features\u003csup\u003e34\u003c/sup\u003e, with initial tests on Toshiba devices showing 12.5% specificity improvement (from 71.4% to 83.9%, n=28 cases)\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4. Conclusion\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIANet provides a non-invasive, interpretable, and scalable tool for IgAN detection. Its high performance (internal AUC=0.975, external AUC=0.9275), speed (0.04 s per case), and validation against histopathology/biomarkers make it a strong candidate for clinical adoption. By enabling early, standardized diagnosis, IANet has the potential to optimizing clinical management pathways for IgAN by integrating standardized imaging metrics into patient evaluation.\u0026nbsp;\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e\u003cstrong\u003e1. Study Design and Participants\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis multi-centre retrospective case-control study included four institutions in China: Primary centre: Sichuan Provincial People\u0026rsquo;s Hospital (Chengdu) \u0026ndash; model training (n=1,256) and internal validation (n=538). External centres:\u0026nbsp;The First Affiliated Hospital of Chongqing Medical University (Chongqing), Sichuan Integrative Medicine Hospital (Chengdu), The First Affiliated Hospital of Ningbo University (Ningbo) \u0026ndash; external validation (n=218 total). Inclusion criteria (IgAN group): \u0026ge;18 years old; biopsy-confirmed IgAN (per Oxford MEST-C classification\u003csup\u003e8\u003c/sup\u003e); renal ultrasound performed within 30 days of biopsy (to ensure structural-pathological consistency); complete clinical data (eGFR, serum creatinine, proteinuria). Inclusion criteria (control group): Healthy volunteers from hospital health management centres; normal renal function (eGFR \u0026ge;90 mL/min/1.73m\u0026sup2;); no abnormal urinalysis (no hematuria/proteinuria); no history of kidney disease, hypertension, or diabetes. Exclusion criteria: Incomplete clinical/imaging data; poor-quality ultrasound images (motion artifacts, insufficient kidney visualization); secondary IgA deposition (e.g., lupus nephritis, Henoch-Sch\u0026ouml;nlein purpura); coexisting autoimmune diseases (e.g., rheumatoid arthritis). \u0026nbsp;The study was approved by the Institutional Review Board of Sichuan Provincial People\u0026rsquo;s Hospital (No. 2025-294). Informed consent was waived due to the retrospective design, with all data de-identified per HIPAA guidelines. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2. Data Acquisition and Preprocessing\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eUltrasound protocol\u003c/strong\u003e: Uniform acquisition using curvilinear array transducers (2\u0026ndash;5 MHz) across 6 manufacturers (GE, Philips, Samsung, Mindray, Toshiba, Canon). For each participant, a board-certified sonographer acquired 3\u0026ndash;5 maximal longitudinal sections of each kidney (left/right); the section with the clearest corticomedullary junction was selected (final dataset: 3,588 images; 1,794 participants \u0026times; 2 kidneys). \u0026nbsp; \u003cstrong\u003eImage quality assessment\u003c/strong\u003e: A radiologist (10-yr exp) scored images on a 5-point scale (1=uninterpretable, 5=excellent visualization); only images with scores \u0026ge;3 were included. \u003cstrong\u003eImage Preprocessing:\u0026nbsp;\u003c/strong\u003eAll preprocessing was implemented in Python 3.9 using OpenCV 4.8.0 and PyTorch 2.0.1. To standardize inputs across heterogeneous scanners, all images were resized to 512\u0026times;512 pixels using bicubic interpolation. This resolution was chosen based on pre-experiments balancing three criteria: optimal retention of the corticomedullary junction features (Dice Similarity Coefficient = 0.91), computational efficiency (0.04 s inference time per image on an NVIDIA Tesla 4070 GPU), and compatibility with both high- and low-resolution native scanner outputs. \u003cstrong\u003eData Augmentation:\u003c/strong\u003e To improve model generalizability, a series of augmentations were applied only to the training set, with all randomness controlled by a fixed seed (seed=42) for reproducibility. This included: (1) Addition of Gaussian noise (\u0026sigma;\u0026nbsp;= 0.01) to simulate inherent ultrasound signal variability. (2) Random rotation within a range of\u0026nbsp;\u0026plusmn;15\u0026deg;. (3) A centered crop retaining 85% of the original area to focus the model on the renal parenchyma. ()4 Horizontal flipping with a 50% probability, applied strategically to kidney pairs to maintain anatomically meaningful bilateral symmetry for the BRSA module, with careful preservation of original side labels.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3. Model Architecture\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe IANet architecture integrates five synergistic modules (Figure 4, labeled signal flow), comprising:\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eShared Feature Extraction Module\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eBackbone: ResNet-18 pretrained on ImageNet\u003csup\u003e35\u003c/sup\u003e\u0026mdash;selected for its balance of performance and efficiency. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eProcessing pipeline: 7\u0026times;7 convolution (stride=2, 64 filters) \u0026rarr; batch normalization\u003csup\u003e36\u003c/sup\u003e \u0026rarr; ReLU activation \u0026rarr; max-pooling (3\u0026times;3, stride=2) \u0026rarr; four residual blocks (layer1: 64 filters; layer2: 128 filters; layer3: 256 filters; layer4: 512 filters) \u0026rarr; 512-channel feature maps (spatial resolution: 16\u0026times;16, 1/32 of input).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eIndividual Feature Branches\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTwo symmetric, independent branches (one for left kidney, one for right kidney) to extract kidney-specific features. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eEach branch includes: Channel-wise attention (1\u0026times;1 convolution + sigmoid activation) to recalibrate feature importance \u0026rarr; 1\u0026times;1 convolution to compress features from 512\u0026rarr;256 channels (reducing computational overhead). \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCross-Interaction Module\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eMulti-head attention (4 heads) to model bidirectional feature dependencies between left and right kidneys. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eProcessing steps were flatten 256\u0026times;16\u0026times;16 feature maps into 256\u0026times;256 sequences. Compute attention weights for left\u0026rarr;right and right\u0026rarr;left interactions (modeling pathological asymmetries). \u0026nbsp;Reconstruct spatial dimensions (256\u0026times;16\u0026times;16) to preserve anatomical localization. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEnhanced Difference Module\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThree-stage pipeline to amplify discriminative inter-kidney disparities: Absolute differencing: Compute pixel-wise feature deviations (|f_left \u0026minus; f_right|) to capture baseline asymmetries. Convolutional enhancement: 3\u0026times;3 convolution (512 filters) + batch normalization + GELU activation\u003csup\u003e37\u003c/sup\u003e to extract multi-scale differential patterns. Joint spatial-channel attention: Channel weighting (reduction MLP\u003csup\u003e38\u003c/sup\u003e, hidden dimension=128) + 7\u0026times;7 convolution (spatial focusing) to amplify disease-specific biomarkers (e.g., cortical echogenicity differences). \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMulti-Dimensional Fusion and Classification\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFeature integration: (1) Global average pooling: Compress cross-interaction features (256\u0026times;16\u0026times;16) into 256-dimensional vectors; compress difference features (512\u0026times;16\u0026times;16) into 512-dimensional vectors. (2) Concatenation: Combine vectors into 768-dimensional unified representations (correction from prior 1024D, reflecting actual feature dimensions). \u0026nbsp;(3) Projection: Layer normalization \u0026rarr; SiLU Sigmoid-weighted linear units for neural network function approximation in reinforcement learning \u0026rarr; dropout (rate=0.3) to prevent overfitting. \u0026nbsp;Classification: Linear layer (1 output) + sigmoid activation to generate IgAN probability scores (0\u0026ndash;1). \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4. Training Protocol\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe model was trained with rigorous protocols to ensure reproducibility and performance. All modules were initialized with a fixed random seed (seed=42). The ResNet-18 backbone was initialized with weights pretrained on ImageNet, while the remaining components\u0026mdash;including the convolutional layers in the Individual Feature Branches, Enhanced Difference module, and Spatial-Channel Attention module\u0026mdash;were initialized using He normal initialization\u003csup\u003e35\u003c/sup\u003e. The multi-head attention layer within the Cross-Interaction module was initialized with Xavier uniform initialization, and all bias terms were set to zero.\u003c/p\u003e\n\u003cp\u003eHyperparameters were optimized via 5-fold cross-validation on the training cohort (n=1,256) using a grid search strategy, with validation loss as the primary selection metric. The comprehensive grid search results for key hyperparameters (optimizer, learning rate, batch size, etc.) are detailed in Supplementary Table 7. The final model was trained for 300 epochs using the Adam optimizer with a learning rate of 0.001, which provided superior convergence speed and lower validation loss compared to Stochastic Gradient Descent (SGD). A batch size of 32 was selected to maximize the memory utilization of the NVIDIA Tesla 4070 GPU. To mitigate overfitting, L2 weight decay (1\u0026times;10⁻⁴) and dropout (rate=0.3) were applied. Gradient clipping (max norm=1.0) was employed to prevent gradient explosion. The loss function was binary cross-entropy, with class weights set to 1:1.08 to balance the IgAN and control classes\u003csup\u003e39\u003c/sup\u003e. Early stopping was triggered if the validation loss failed to improve for 15 consecutive epochs, and the model checkpoint with the lowest validation loss was retained.\u003c/p\u003e\n\u003cp\u003eTo quantitatively evaluate the contribution of the proposed Bilateral Renal Symmetry-Aware (BRSA) components, an ablation study was conducted on the internal validation cohort (n=538). Four model variants\u0026mdash;all sharing the same ResNet-18 backbone and training hyperparameters\u0026mdash;were compared: (1) a baseline model without any BRSA components; (2) a model with only the Cross-Interaction module; (3) a model with only the Enhanced Difference module; and (4) the full IANet incorporating all BRSA submodules.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e5. Histopathological and Biomarker Correlation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePathologist review\u003c/strong\u003e: Two renal pathologists (\u0026gt;15-yr exp) independently reviewed Grad-CAM heatmaps and corresponding biopsy slides in a blinded manner (no knowledge of model outputs). \u0026nbsp;\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCorrelation methods:\u0026nbsp;\u003c/strong\u003eTo quantify the heterogeneity of IANet\u0026apos;s attention, the renal parenchyma was first manually segmented on each ultrasound image by a radiologist (10-yr exp) blinded to the model and pathology results. This defined the parenchymal region of interest (ROI). Then, gray level co-occurrence matrix-angular second moment (GLCM-ASM) was adopted to characterize parenchymal texture uniformity (and indirectly reflect attention heterogeneity, with lower ASM indicating higher texture/attention heterogeneity). The GLCM-ASM calculation for irregular ROIs included mask preprocessing, valid pixel pair screening (4 directions, d=1 pixel), symmetric and normalized GLCM construction, and arithmetic averaging of 4-direction ASM values. The mesangial proliferation (M1) subgroup was excluded due to insufficient sample size (n=18, 5.0%).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e6. Statistical Analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll statistical analyses were performed using Python 3.9 (with SciPy 1.10.1, scikit-learn 1.2.2, and Statsmodels 0.13.5 packages) and SPSS 29.0, with visualization conducted using Matplotlib 3.7.1. Model performance was evaluated primarily by the area under the receiver operating characteristic curve (AUC). The 95% confidence intervals (CIs) for AUC values were computed using the DeLong method, while the Wilson score interval method was employed for proportions such as accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV).\u003c/p\u003e\n\u003cp\u003eThe sample size for the internal validation cohort was calculated to be 803 participants to achieve 95% accuracy with a\u0026nbsp;\u0026plusmn;1.5% precision. After data quality filtering, 538 participants were ultimately included. For the external validation cohort, a sample size of 198 was calculated to be necessary to detect an AUC difference of 0.10 compared to conventional ultrasound (AUC=0.82), with a two-sided\u0026nbsp;\u0026alpha;\u0026nbsp;of 0.05 and 90% power; 218 participants were enrolled, achieving a power of 92.3%. To mitigate potential class imbalance, class weights were set at 1:1.08 based on the prevalence of IgAN (47.9%) in the training cohort.\u003c/p\u003e\n\u003cp\u003eFor statistical comparisons, categorical variables were analyzed using the Chi-square test or Fisher\u0026apos;s exact test when cell counts were less than five. Continuous variables that were normally distributed were compared using the independent samples t-test, while the Mann-Whitney U test was applied for non-normally distributed data. Comparisons of diagnostic performance among multiple radiologists were conducted using Cochran\u0026apos;s Q test, followed by post-hoc McNemar tests with a Bonferroni-adjusted significance level (\u0026alpha; = 0.0083). Pairwise comparisons of AUC values were performed using DeLong\u0026apos;s test. Inter-rater agreement was assessed using Fleiss\u0026apos; \u0026kappa;, interpreted as follows: \u0026lt;0.2, poor; 0.2\u0026ndash;0.4, fair; 0.4\u0026ndash;0.6, moderate; and \u0026gt;0.6, substantial.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAuthor Contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eYanyan Wu: Conceptualization, methodology, writing—original draft. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eJinqi Gong: Software, model training, writing—original draft. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eMingRui Fu: Formal analysis, data curation. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eZiHan Tang: Visualization. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eYunhao Luo, XueYing Chen, Yong Jin: Clinical data collection (ultrasound images, biopsy reports). \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eChengyuan Ye: Data curation, project administration.\u003c/p\u003e\n\u003cp\u003eXinYi Hu: Formal analysis, validation.\u003c/p\u003e\n\u003cp\u003eDong Wang, Ping Zhang: Supervision, writing—review \u0026amp; editing (validated statistical analyses). \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAll authors discussed the results and contributed to the final manuscript. \u0026nbsp;Yanyan Wu and Jinqi Gong contributed equally to this work as co-first authors. Dong Wang and Ping Zhang jointly supervised the project and served as co-corresponding authors.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCode availability\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe code used to train and evaluate the model is available on GitHub (https://github.com/wyydoctor/IgAN_model.git).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting Interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe thank Professor Guisen Li (Department of Nephrology, Sichuan Provincial People’s Hospital) for clinical insights into IgAN diagnosis and management, and the three renal pathologists (Drs. Li, Wang, and Zhang) for blinded histopathological review of biopsy slides (critical for Table 5 analyses). \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics Statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe protocol of this study has been approved by the Institutional Review Board of Sichuan Provincial People’s Hospital (No. 2025-294). Due to the retrospective nature of the study, it meets the requirements for waiver of informed consent. We confirm that we have strictly adhered to the original protocol for data processing authorized by the IRB\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. \u0026nbsp;\u0026nbsp;\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eLi LS, Liu ZH. Epidemiologic data of renal diseases from a single unit in China: analysis based on 13,519 renal biopsies. \u003cem\u003eKidney Int\u003c/em\u003e 66, 920\u0026ndash;923 (2004).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFilippone EJ, Gulati R, Farber JL. Contemporary review of IgA nephropathy. \u003cem\u003eFront Immunol\u003c/em\u003e 15, 1436923 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePattrapornpisut P, Avila-Casado C, Reich HN. IgA Nephropathy: Core Curriculum 2021. \u003cem\u003eAm J Kidney Dis\u003c/em\u003e 78, 429\u0026ndash;441 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWyatt RJ, Julian BA. IgA nephropathy. \u003cem\u003eN Engl J Med\u003c/em\u003e 368, 2402\u0026ndash;2414 (2013).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCheng HT, Xu X, Lim PS, Hung KY. Worldwide Epidemiology of Diabetes-Related End-Stage Renal Disease, 2000\u0026ndash;2015. \u003cem\u003eDiabetes Care\u003c/em\u003e 44, 89\u0026ndash;97 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBarbour SJ, \u003cem\u003eet al.\u003c/em\u003e Evaluating a New International Risk-Prediction Tool in IgA Nephropathy. \u003cem\u003eJAMA Intern Med\u003c/em\u003e 179, 942\u0026ndash;952 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSuzuki H, \u003cem\u003eet al.\u003c/em\u003e Aberrantly glycosylated IgA1 in IgA nephropathy patients is recognized by IgG antibodies with restricted heterogeneity. \u003cem\u003eJ Clin Invest\u003c/em\u003e 119, 1668\u0026ndash;1677 (2009).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRoberts IS, \u003cem\u003eet al.\u003c/em\u003e The Oxford classification of IgA nephropathy: pathology definitions, correlations, and reproducibility. \u003cem\u003eKidney Int\u003c/em\u003e 76, 546\u0026ndash;556 (2009).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTrimarchi H, \u003cem\u003eet al.\u003c/em\u003e Oxford Classification of IgA nephropathy 2016: an update from the IgA Nephropathy Classification Working Group. \u003cem\u003eKidney Int\u003c/em\u003e 91, 1014\u0026ndash;1021 (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHogan JJ, Mocanu M, Berns JS. The Native Kidney Biopsy: Update and Evidence for Best Practice. \u003cem\u003eClin J Am Soc Nephrol\u003c/em\u003e 11, 354\u0026ndash;362 (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKamano C, \u003cem\u003eet al.\u003c/em\u003e A cross-sectional study in patients with IgA nephropathy of correlations between clinical data and pathological findings at the time of renal biopsy: a Japanese prospective cohort study. \u003cem\u003eClin Exp Nephrol\u003c/em\u003e 25, 509\u0026ndash;521 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMoghazi S, \u003cem\u003eet al.\u003c/em\u003e Correlation of renal histopathology with sonographic findings. \u003cem\u003eKidney Int\u003c/em\u003e 67, 1515\u0026ndash;1520 (2005).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLib\u0026oacute;rio AB, de Oliveira Neves FM, Torres de Melo CB, Leite TT, de Almeida Leit\u0026atilde;o R. Quantitative Renal Echogenicity as a Tool for Diagnosis of Advanced Chronic Kidney Disease in Patients with Glomerulopathies and no Liver Disease. Kidney Blood Press Res 42, 708\u0026ndash;716 (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen Z, \u003cem\u003eet al.\u003c/em\u003e Development and Deployment of a Novel Diagnostic Tool Based on Conventional Ultrasound for Fibrosis Assessment in Chronic Kidney Disease. \u003cem\u003eAcad Radiol\u003c/em\u003e 30 Suppl 1, S295-s304 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eQin X, Xia L, Ma Q, Cheng D, Zhang C. Development of a novel combined nomogram model integrating deep learning radiomics to diagnose IgA nephropathy clinically. \u003cem\u003eRen Fail\u003c/em\u003e 45, 2271104 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFeretzakis G, \u003cem\u003eet al.\u003c/em\u003e Emerging Trends in AI and Radiomics for Bladder, Kidney, and Prostate Cancer: A Critical Review. \u003cem\u003eCancers (Basel)\u003c/em\u003e 16, (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePeng T, \u003cem\u003eet al.\u003c/em\u003e A multi-center study of ultrasound images using a fully automated segmentation architecture. \u003cem\u003ePattern Recognition\u003c/em\u003e 145, (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLu WW, Zhang D, Ni XJ. A Review of the Role of Ultrasound Radiomics and Its Application and Limitations in the Investigation of Thyroid Disease. \u003cem\u003eMed Sci Monit\u003c/em\u003e 28, e937738 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLundberg SM, Lee SI. A Unified Approach to Interpreting Model Predictions. In: 31st Annual Conference on Neural Information Processing Systems (NIPS)) (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuang X, Hu Y, Zhang Y, Zhou Q. Two-Dimensional Ultrasound-Based Radiomics Nomogram for Diabetic Kidney Disease: A Pilot Study. Int J Gen Med 17, 1877\u0026ndash;1885 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHasegawa K, \u003cem\u003eet al.\u003c/em\u003e Inflammation-associated molecules in the glomerular-endothelium in mild IgA-nephropathy patients identified by single-cell and spatial transcriptome. \u003cem\u003eCommun Biol\u003c/em\u003e 8, 915 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKrizhevsky A, Sutskever I, Hinton GE. ImageNet classification with deep convolutional neural networks. \u003cem\u003eCommun ACM\u003c/em\u003e 60, 84\u0026ndash;90 (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuang X, Hu Y, Zhang Y, Zhou Q. Two-Dimensional Ultrasound-Based Radiomics Nomogram for Diabetic Kidney Disease: A Pilot Study. \u003cem\u003eInt J Gen Med\u003c/em\u003e 17, 1877\u0026ndash;1885 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang L, \u003cem\u003eet al.\u003c/em\u003e Preliminary study on the application of renal ultrasonography radiomics in the classification of glomerulopathy. \u003cem\u003eBMC Med Imaging\u003c/em\u003e 21, 115 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKuo CC, \u003cem\u003eet al.\u003c/em\u003e Automation of the kidney function prediction and classification through ultrasound-based kidney imaging using deep learning. \u003cem\u003eNPJ Digit Med\u003c/em\u003e 2, 29 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTjoa E, Guan C. A Survey on Explainable Artificial Intelligence (XAI): Toward Medical XAI. IEEE Trans Neural Netw Learn Syst 32, 4793\u0026ndash;4813 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRudin C. Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable Models Instead. \u003cem\u003eNat Mach Intell\u003c/em\u003e 1, 206\u0026ndash;215 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBellur SS, Troyanov S, Vorobyeva O, Coppo R, Roberts ISD. Evidence from the large VALIGA cohort validates the subclassification of focal segmental glomerulosclerosis in IgA nephropathy. \u003cem\u003eKidney Int\u003c/em\u003e 105, 1279\u0026ndash;1290 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRippy JR, Singh M, Aglyamov SR, Larin KV. Ultrasound Shear Wave Elastography and Transient Optical Coherence Elastography: Side-by-Side Comparison of Repeatability and Accuracy. \u003cem\u003eIEEE Open Journal of Engineering in Medicine and Biology\u003c/em\u003e 2, 179\u0026ndash;186 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChang TW, \u003cem\u003eet al.\u003c/em\u003e Artificial intelligence for predicting interstitial fibrosis and tubular atrophy using diagnostic ultrasound imaging and biomarkers. \u003cem\u003eBMJ Health Care Inform\u003c/em\u003e 32, (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhu W, \u003cem\u003eet al.\u003c/em\u003e A clinically applicable CKD diagnostic model derived from sound touch viscosity ultrasound and LASSO regression. \u003cem\u003eFront Bioeng Biotechnol\u003c/em\u003e 13, 1651500 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang X, \u003cem\u003eet al.\u003c/em\u003e The impact of the combat method on radiomics feature compensation and analysis of scanners from different manufacturers. \u003cem\u003eBMC Med Imaging\u003c/em\u003e 24, 137 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKhaked AA, Oishi N, Roggen D, Lago P. In Shift and In Variance: Assessing the Robustness of HAR Deep Learning Models Against Variability. \u003cem\u003eSensors (Basel)\u003c/em\u003e 25, (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhou SK, et al. A review of deep learning in medical imaging: Imaging traits, technology trends, case studies with progress highlights, and future promises. Proc IEEE Inst Electr Electron Eng 109, 820\u0026ndash;838 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHe KM, Zhang XY, Ren SQ, Sun J, Ieee. Deep Residual Learning for Image Recognition. In: 2016 \u003cem\u003eIEEE Conference on Computer Vision and Pattern\u003c/em\u003e Recognition \u003cem\u003e(CVPR)\u003c/em\u003e) (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIoffe S, Szegedy C. Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. In: \u003cem\u003e32nd International Conference on Machine Learning\u003c/em\u003e) (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMatias R, Jorge G-S, Victor AJ, Adrian W. Zorro: A Flexible and Differentiable Parametric Family of Activation Functions That Extends ReLU and GELU.) (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eQifeng Z, Barry YC, Nelson M, Andreas S. On using MLP features in LVCSR.) (2004).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eQian Q, Jin R, Yi J, Zhang L, Zhu S. Efficient distance metric learning by adaptive sampling and mini-batch stochastic gradient descent (SGD). \u003cem\u003eSpringer US\u003c/em\u003e, (2015).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"},{"header":"Tables","content":"\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"703\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"5\" valign=\"top\" style=\"width: 683px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTable 1\u003c/strong\u003e Baseline characteristics.\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 20px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 192px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCharacteristics\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTraining Cohort\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 155px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTest Cohort\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 184px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eExternal Validation Cohort\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 20px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 192px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003eIgAN (n=601) \u0026nbsp; Controls (n=655) \u0026nbsp;\u003cem\u003ep\u003c/em\u003e-value\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 155px;\"\u003e\n \u003cp\u003eIgAN (n=257) \u0026nbsp; Controls (n=281) \u0026nbsp;\u003cem\u003ep\u003c/em\u003e-value\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 184px;\"\u003e\n \u003cp\u003eIgAN (n=105) \u0026nbsp; Controls (n=113) \u0026nbsp; \u0026nbsp;\u003cem\u003ep\u003c/em\u003e-value\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 20px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 192px;\"\u003e\n \u003cp\u003eAge (years, mean \u0026plusmn; SD)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e33.21 \u0026plusmn; 11.70 \u0026nbsp; \u0026nbsp;34.29 \u0026plusmn; 12.43 \u0026nbsp; \u0026nbsp; 0.08\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 155px;\"\u003e\n \u003cp\u003e34.04\u0026plusmn; 12.15 \u0026nbsp; \u0026nbsp;32.91 \u0026plusmn; 12.06 \u0026nbsp; \u0026nbsp; 0.28\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e34.50 \u0026plusmn; 12.20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 136px;\"\u003e\n \u003cp\u003e33.80 \u0026plusmn; 11.90 \u0026nbsp; \u0026nbsp; 0.12\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 192px;\"\u003e\n \u003cp\u003eGender (Male/Female)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e327/27 \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;4336/319 \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;0.15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 155px;\"\u003e\n \u003cp\u003e142/115 \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;162/119 \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;0.58\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e50/55\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 136px;\"\u003e\n \u003cp\u003e60/53 \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; 0.42\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 192px;\"\u003e\n \u003cp\u003eSerum Creatinine (\u0026mu;mol/L)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e134.07 \u0026plusmn; 18.07 80.10 \u0026plusmn; 11.43 ＜0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 155px;\"\u003e\n \u003cp\u003e131.43\u0026plusmn; 7.35 81.30 \u0026plusmn; 11.02 ＜0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e132.55 \u0026plusmn; 17.55\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 136px;\"\u003e\n \u003cp\u003e80.55 \u0026plusmn; 11.25 ＜0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 192px;\"\u003e\n \u003cp\u003eeGFR (mL/min/1.73m\u0026sup2;)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e72.74 \u0026plusmn; 15.43 103.31 \u0026plusmn; 9.80 ＜0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 155px;\"\u003e\n \u003cp\u003e70.51\u0026plusmn; 13.95 102.66 \u0026plusmn; 11.98 ＜0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e70.85 \u0026plusmn; 14.55\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 136px;\"\u003e\n \u003cp\u003e102.05 \u0026plusmn; 10.04 ＜0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 192px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 152px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 155px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 68px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 20px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"100%\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"7\" valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTable 2 Diagnostic performance of baseline models versus IANet variants Performance metrics include accuracy, AUC, TPR, and FPR with 95% confidence intervals\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eModel\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAccuracy (95% CI)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAUC (95% CI)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTPR (95% CI)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eFPR (95% CI)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e\u003cstrong\u003ePPV (95% CI)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eNPV (95% CI)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eResNet-18 Baseline\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e84.47%\u003c/p\u003e\n \u003cp\u003e(82.35\u0026ndash;86.32%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e0.892\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(0.871\u0026ndash;0.910)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e81.3%\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(78.5\u0026ndash;83.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e15.6%\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(13.2\u0026ndash;17.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e82.9%\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(78.2\u0026ndash;87.1%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e83.2%\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(78.5\u0026ndash;87.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eIANet (ResNet-18)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e96.73%\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(95.81\u0026ndash;97.49%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e0.975\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(0.962\u0026ndash;0.984)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e95.2%\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(93.1\u0026ndash;96.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e3.1%\u003c/p\u003e\n \u003cp\u003e(2.3\u0026ndash;4.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e96.8%\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(94.1\u0026ndash;98.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e95.8%\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(92.6\u0026ndash;97.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eResNet-50 Baseline\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e88.62%\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(86.85\u0026ndash;90.18%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e0.921\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(0.903\u0026ndash;0.936)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e85.7%\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(83.2\u0026ndash;87.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e11.2%\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(9.5\u0026ndash;13.1%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e87.7%\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(83.5\u0026ndash;91.1%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e87.1%\u003c/p\u003e\n \u003cp\u003e(82.8\u0026ndash;90.6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eIANet (ResNet-50)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e95.82%\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(94.73\u0026ndash;96.69%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e0.968\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(0.953\u0026ndash;0.979)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e93.8%\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(91.5\u0026ndash;95.6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e4.5%\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(3.6\u0026ndash;5.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e94.9%\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(91.5\u0026ndash;97.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e94.4%\u003c/p\u003e\n \u003cp\u003e(90.8\u0026ndash;96.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDenseNet-121 Baseline\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e86.35%\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;(84.42\u0026ndash;88.09%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e0.905\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(0.884\u0026ndash;0.923)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e83.1%\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(80.4\u0026ndash;85.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e13.5%\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(11.7\u0026ndash;15.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e84.9%\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(80.4\u0026ndash;88.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e84.7%\u003c/p\u003e\n \u003cp\u003e(79.9\u0026ndash;88.6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eIANet (DenseNet-121)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e94.95%\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(93.68\u0026ndash;95.98%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e0.961\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(0.945\u0026ndash;0.973)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e92.1%\u003c/p\u003e\n \u003cp\u003e(89.7\u0026ndash;94.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e5.8%\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(4.9\u0026ndash;6.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e93.7%\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(90.0\u0026ndash;96.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e93.0%\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(89.1\u0026ndash;95.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eInceptionV3 Baseline\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e85.12%\u003c/p\u003e\n \u003cp\u003e(83.15\u0026ndash;86.89%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e0.898\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(0.877\u0026ndash;0.916)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e82.4%\u003c/p\u003e\n \u003cp\u003e(79.6\u0026ndash;84.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e14.3%\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(12.5\u0026ndash;16.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e84.1%\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(79.7\u0026ndash;87.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e84.3%\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(79.8\u0026ndash;87.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eIANet (InceptionV3)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15px;\"\u003e\n \u003cp\u003e93.27%\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(91.92\u0026ndash;94.38%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e0.952\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(0.935\u0026ndash;0.965)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e89.5%\u003c/p\u003e\n \u003cp\u003e(87.0\u0026ndash;91.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e7.3%\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(6.2\u0026ndash;8.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e91.6%\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(87.8\u0026ndash;94.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e90.6%\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e(86.7\u0026ndash;93.6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" align=\"\" width=\"660\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"5\" valign=\"top\" style=\"width: 660px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTable 3 Diagnostic performance of IANet in the multi-center external validation cohort\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 135px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMetric Category\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 202px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMetric Name\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAbbreviation\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 84px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eValue\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 157px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e95% Confidence Interval (95% CI)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 135px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCohort Characteristics\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 202px;\"\u003e\n \u003cp\u003eTotal Samples\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 84px;\"\u003e\n \u003cp\u003e218\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 157px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 135px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 202px;\"\u003e\n \u003cp\u003eBiopsy-Confirmed IgAN Cases\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 84px;\"\u003e\n \u003cp\u003e105\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 157px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 135px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 202px;\"\u003e\n \u003cp\u003eNon-IgAN Controls (including healthy/other CKD)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 84px;\"\u003e\n \u003cp\u003e113\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 157px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 135px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCore Diagnostic Efficacy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 202px;\"\u003e\n \u003cp\u003eArea Under ROC Curve\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003eAUC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 84px;\"\u003e\n \u003cp\u003e0.9275\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 157px;\"\u003e\n \u003cp\u003e0.8912\u0026ndash;0.9548\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 135px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 202px;\"\u003e\n \u003cp\u003eOverall Accuracy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003eAcc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 84px;\"\u003e\n \u003cp\u003e0.8624\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 157px;\"\u003e\n \u003cp\u003e0.8115\u0026ndash;0.9037\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 135px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eConfusion Matrix\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 202px;\"\u003e\n \u003cp\u003eTrue Positive (IgAN correctly classified)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003eTP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 84px;\"\u003e\n \u003cp\u003e101\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 157px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 135px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 202px;\"\u003e\n \u003cp\u003eTrue Negative (Non-IgAN correctly classified)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003eTN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 84px;\"\u003e\n \u003cp\u003e87\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 157px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 135px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 202px;\"\u003e\n \u003cp\u003eFalse Positive (Non-IgAN misclassified as IgAN)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003eFP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 84px;\"\u003e\n \u003cp\u003e26\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 157px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 135px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 202px;\"\u003e\n \u003cp\u003eFalse Negative (IgAN misclassified as Non-IgAN)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003eFN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 84px;\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 157px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 135px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eClinical Risk Metrics\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 202px;\"\u003e\n \u003cp\u003eSensitivity (TPR, avoiding missed diagnosis)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003eSen\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 84px;\"\u003e\n \u003cp\u003e0.9619\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 157px;\"\u003e\n \u003cp\u003e0.9157\u0026ndash;0.9864\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 135px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 202px;\"\u003e\n \u003cp\u003eSpecificity (1-FPR, avoiding misdiagnosis)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003eSpe\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 84px;\"\u003e\n \u003cp\u003e0.7700\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 157px;\"\u003e\n \u003cp\u003e0.6853\u0026ndash;0.8398\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 135px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDiagnostic\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 202px;\"\u003e\n \u003cp\u003eTotal Diagnostic Time (for 218 samples)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 84px;\"\u003e\n \u003cp\u003e8.85 s\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 157px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"624\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"5\" valign=\"top\" style=\"width: 624px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTable\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;4\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003eDiagnostic Performance Comparison\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eEvaluator\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAccuracy (95% CI)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 122px;\"\u003e\n \u003cp\u003e\u003cstrong\u003ePPV (95% CI)\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eNPV (95% CI)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e\u003cstrong\u003ediagnostic\u0026nbsp;Time (s)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eIANet\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e98.0% (95.3-100.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 122px;\"\u003e\n \u003cp\u003e100.0% (91.2-100.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e96.7% (88.7-99.6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRadiologist X. G\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e82.0% (74.4-89.6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 122px;\"\u003e\n \u003cp\u003e80.0% (64.3-90.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e83.3% (71.5-91.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e2040\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRadiologist J. S\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e74.0% (65.5-82.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 122px;\"\u003e\n \u003cp\u003e94.4% (72.7-99.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e69.5% (58.7-78.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e1680\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRadiologist F. H\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e71.0% (62.2-79.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 122px;\"\u003e\n \u003cp\u003e84.2% (60.4-96.6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e67.9% (57.1-77.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e2160\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" align=\"\" width=\"76%\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"6\" valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTable 5\u0026nbsp;\u003c/strong\u003eCorrelation Analysis of IANet\u0026rsquo;s Attention Patterns with Histopathological Features\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 28px;\"\u003e\n \u003cp\u003ePathological/Clinical Feature\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003eCorrelation Measure\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e95% CI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u003cem\u003ep\u003c/em\u003e-value\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003eSample Size (n)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 21px;\"\u003e\n \u003cp\u003eAnalysis Method\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 28px;\"\u003e\n \u003cp\u003eHistopathological Correlation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 21px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 28px;\"\u003e\n \u003cp\u003eMesangial proliferation (M1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e18(5.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 21px;\"\u003e\n \u003cp\u003eexcluded,\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 28px;\"\u003e\n \u003cp\u003eSegmental sclerosis (S1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003eOR = 4.28\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e2.95\u0026ndash;6.21\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e198(54.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 21px;\"\u003e\n \u003cp\u003eMultivariate logistic regression\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 28px;\"\u003e\n \u003cp\u003eEndothelial proliferation (E1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003eOR = 2.15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e1.32\u0026ndash;3.51\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e0.002\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e97(26.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 21px;\"\u003e\n \u003cp\u003eMultivariate logistic regression\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 28px;\"\u003e\n \u003cp\u003eCrescent formation (C1\u0026amp;C2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003eOR = 1.83\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e0.97\u0026ndash;3.45\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e0.063\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e114(31.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 21px;\"\u003e\n \u003cp\u003eMultivariate logistic regression\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 28px;\"\u003e\n \u003cp\u003eInterstitial fibrosis/tubular atrophy (T1\u0026amp;T2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003eOR = 3.20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e2.25 \u0026ndash; 4.55\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e129(35.6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 113px;\"\u003e\n \u003cp\u003eMultivariate logistic regression\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 28px;\"\u003e\n \u003cp\u003eImaging-Pathology Quantitative Correlation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 21px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 28px;\"\u003e\n \u003cp\u003eCortical echogenicity increase\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003er = 0.69\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e0.59\u0026ndash;0.77\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e87\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 21px;\"\u003e\n \u003cp\u003ePearson correlation\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eNotes:\u003c/p\u003e\n\u003col\u003e\n \u003cli\u003eThe independent variable for all logistic regression models was renal parenchyma gray level co-occurrence matrix-angular second moment (GLCM-ASM) (lower GLCM-ASM values indicate higher parenchymal structural disorganization and corresponding model attention heterogeneity).\u003c/li\u003e\n \u003cli\u003eAll logistic regression models were adjusted for age and gender.\u003c/li\u003e\n \u003cli\u003eAbbreviations: OR = Odds Ratio; CI = Confidence Interval.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"IgA Nephropathy, IANet, Bilateral Renal Symmetry-Aware Analysis (BRSA), Ultrasound Diagnosis, Explainable AI, Histopathological Correlation ","lastPublishedDoi":"10.21203/rs.3.rs-8425698/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8425698/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eImmunoglobulin A nephropathy (IgAN), the most prevalent primary glomerulonephritis globally, lacks reliable non-invasive diagnostics. Current diagnosis depends on invasive renal biopsy (bleeding/pain risks), and conventional ultrasound misses early microstructural lesions. Here, we present IANet, an interpretable deep learning framework using Bilateral Renal Symmetry-Aware Analysis (BRSA) to amplify subtle IgAN-related sonographic changes.\u003c/p\u003e \u003cp\u003eIn a 4-centre retrospective study of 2,012 participants, IANet was trained (n\u0026thinsp;=\u0026thinsp;1,256), internally (n\u0026thinsp;=\u0026thinsp;538) and externally (n\u0026thinsp;=\u0026thinsp;218) validated. It achieved 96.7% accuracy (95% CI: 95.8\u0026ndash;97.5%) and 0.975 AUC (95% CI: 0.962\u0026ndash;0.984) internally, with 80.1% lower false positive rate (3.1% vs. 15.6% for ResNet-18). Externally (six ultrasound manufacturers), it maintained 86.2% accuracy (95% CI: 81.2\u0026ndash;90.4%), 0.928 AUC (95% CI: 0.891\u0026ndash;0.955%), 96.2% sensitivity, and 77.0% specificity.\u003c/p\u003e \u003cp\u003eBlinded tests vs. 3 radiologists (10\u0026ndash;15 years\u0026rsquo; experience) showed IANet outperformed humans (98.0% vs. 71.0\u0026ndash;82.0% accuracy, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001) and was \u0026gt;\u0026thinsp;1,000-fold faster (4 s vs. 28\u0026ndash;36 min per 100 cases). Subgroup analysis stratified by the Oxford MEST-C classification demonstrated that IANet achieved its peak diagnostic accuracy for S1 and T2 lesions in IgAN, while a modest performance decline was observed in cases with T0 pathology. Histopathological validation confirmed that IANet's attention heterogeneity was significantly associated with chronic lesions, demonstrating strong predictive value for segmental sclerosis (S1, OR\u0026thinsp;=\u0026thinsp;4.28, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001) and interstitial fibrosis/tubular atrophy (T1\u0026amp;T2, OR\u0026thinsp;=\u0026thinsp;3.20, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001). In a DKD cohort (n\u0026thinsp;=\u0026thinsp;192), it correctly classified 64.1% as non-IgAN.\u003c/p\u003e \u003cp\u003eIANet is a novel tool for non-invasive and dynamic renal function monitoring, thereby facilitating the advancement of precision nephrology.\u003c/p\u003e","manuscriptTitle":"A Deep Learning Framework Based on Ultrasomics for IgA Nephropathy Detection: A Multi-Centre, Multi-Scanner Validation Study","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-01-30 09:00:38","doi":"10.21203/rs.3.rs-8425698/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"9c8eaaf0-0264-43ff-8069-8c2b0d12db71","owner":[],"postedDate":"January 30th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":61991761,"name":"Health sciences/Nephrology/Kidney diseases/Nephritis"},{"id":61991762,"name":"Health sciences/Diseases/Kidney diseases/Chronic kidney disease"}],"tags":[],"updatedAt":"2026-03-26T14:59:30+00:00","versionOfRecord":[],"versionCreatedAt":"2026-01-30 09:00:38","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8425698","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8425698","identity":"rs-8425698","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}