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

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 < 0.001) and was > 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.