Knee osteoarthritis grading without pooled records: ten sites hold feature columns, asynchronous federation reports 99.64% accuracy

Kellgren-Lawrence grading of knee osteoarthritis can proceed without pooling raw records, since institutions exchange model updates rather than patient data. Centralized models previously required gathering structured records, semi-structured files and imaging in one place, bringing high computation cost and single-point leakage risk. On 5,580 knee osteoarthritis cases the scheme reports 99.64% accuracy and 0.998 AUC, above PHMS-Fed at 95.91% and 0.965 and FL-TinyML-DP-HS at 95% and 0.942; for FL-CNN-LSTM-DPET-HS the source gives 92% accuracy and 0.90 AUC, while the same source's stated AUC improvements disagree with its Table 2 values, so no AUC difference is adopted here.

Multimodal data are first fused into one queryable table before feature extraction and federated training. Unified Element Tree Fusion merges matching keys across modalities and integrates feature vectors into Delta tables or a single Spark DataFrame. The data come from Beijing Tsinghua Changgung Hospital structured records, CSV/JSON/XML files and X-ray, MRI, CT and PET imaging, held in an Apache Spark-managed data lake.

Local feature extraction uses a convolutional network with attention modules that emphasize channels and spatial regions tied to joint pathology. A ResNet-34 backbone is followed by Condensed Depthwise Squeeze-Excitation and Multi-Convolution Spatial Attention, weighting channels and spatial locations respectively. The network extracts joint space narrowing, osteophyte formation, cartilage degradation and subchondral bone change from standardized knee images for five-grade Kellgren-Lawrence classification.

Server-side Bulyan aggregation and adaptive differential privacy jointly suppress malicious updates and inference attacks. Bulyan first selects reliable gradients with Multi-Krum, then applies a coordinate-wise trimmed mean; adaptive differential privacy schedules Gaussian noise with exponential decay. Ten simulated institutions compute local gradients on vertically partitioned data, reporting a 1.8% attack success rate, 94% Byzantine resilience, privacy budget 1.5 and a 0.96 privacy-utility index.

The scheme is validated on 5,580 knee osteoarthritis cases covering five Kellgren-Lawrence grades, split 60% training, 20% validation and 20% testing, with ten simulated institutional clients on the federated side. The intended readers are researchers and engineers working on privacy-preserving medical image grading and federated deployment. The authors propose extending the MCSESEN architecture to other musculoskeletal and degenerative conditions and refining tree-structured communication and gradient masking to cut bandwidth needs.

The ten federated institutions are simulated clients rather than a real hospital deployment, and the privacy budget, Byzantine resilience and attack success rate all come from simulated attack settings. The source is internally inconsistent on AUC: Table 2 lists 0.928 for FL-CNN-LSTM-DPET-HS while the related text gives 0.90, and the reported AUC improvements do not match those values, so this report adopts no AUC difference. Performance under differing cross-institution data distributions is not reported, nor are communication measurements under real network conditions.