DU-KAN: A Dual-Domain Enhanced Deep-Learning Framework for Digital Pathology-Based Oxford MEST-C Assessment in Pediatric IgA Nephropathy

Oxford MEST-C assessment of pediatric immunoglobulin A nephropathy (IgAN) relies on pathologists’ interpretation of complex renal biopsy structures. Automated segmentation may support standardized quantification but is challenged by scale variation, weak boundaries, and small lesions. We developed DU-KAN, a stage-adapted U-KAN framework combining a shallow convolution–MaxViT dual-branch encoder with adapted heterogeneous feature fusion and wavelet–prototype skip enhancement. A single-center dataset of 532 periodic acid–Schiff-stained images from 278 children was partitioned at the patient level. On the held-out test set of 55 patients, DU-KAN achieved a mean Dice of 69.58% across eight classes, including background (95% patient-level bootstrap CI, 66.23–71.91), exceeding U-KAN by 3.56 percentage points (95% CI, 1.90–5.26). Foreground mean Dice was 65.42%. On the Synapse abdominal CT benchmark, mean Dice across eight organs was 87.21%, supporting cross-task architectural generalization, rather than external validation in pediatric IgAN. Segmentation outputs were further aggregated into case-level MEST-C-related indicators. In an exploratory 40-case evaluation, complete agreement across the five assessed indicators was 65.00% (26/40), while sensitivity was 37.50% for E (3/8) and 50.00% for T (1/2). The 97.50% agreement for T largely reflected the predominance of negative cases and did not establish reliable positive-case detection. DU-KAN improves the internal segmentation performance and achieves structured pathological quantification, which can provide auxiliary reference for clinical judgment of renal pathologists.