Ilmu Komputer & AI editorial
Open AccessOA2026
OCT-FedSIR: Toward Trustworthy Federated Ophthalmic Learning under Annotation Noise
A reliability-aware spectral framework for federated OCT classification under client-dependent annotation noise and heterogeneous data distributions
Sina Gholami; Abdulmoneam Ali; Tania Haghighi; Rashadul H. Badhon; Behafarin Emam; Sally S. Y. Ong; Atalie C. Thompson; Theodore Leng; Ahmed Arafa; Jennifer I. Lim; Minhaj Nur Alamยท 2026ยท DOI 10.48550/arXiv.2609.14734
The core problem
Federated learning (FL) enables collaborative model development without centralizing patient data, making it attractive for ophthalmic imaging where privacy is paramount. However, annotation reliability at participating institutions cannot always be assumed. In ophthalmic imaging, differences in disease prevalence and class composition can resemble changes caused by corrupted supervision, creating a fundamental challenge: distinguishing genuine heterogeneity from annotation noise. This paper introduces OCT-FedSIR, a reliability-aware spectral framework for federated OCT classification under client-dependent annotation noise and heterogeneous data distributions. The framework addresses the critical need for trustworthy federated learning in medical imaging, where noisy labels can severely degrade model performance and compromise clinical decision-making.
Innovation
The framework was evaluated on the Kermany, University of Illinois Chicago, and Wake Forest datasets under symmetric and structured asymmetric noise and three levels of non-IID heterogeneity. Across 117 experimental conditions, OCT-FedSIR achieved a mean accuracy of 86.73%, compared with 79.94% for RoFL and 78.75% for FedCorr. The framework correctly separated clients with original and corrupted annotations across all evaluated conditions, while the original FedSIR identification procedure was less robust, particularly under asymmetric noise. Spectral relabeling recovered 77.2% of corrupted annotations with 91.3% correction precision and a 3.5% false-correction rate. Retaining corrected clients outperformed spectral pruning by 9.30 percentage points on average. These results demonstrate that annotation noise can often be identified and corrected without discarding informative client data, significantly improving federated learning performance in the presence of label noise.
Federated learning (FL) enables collaborative model development without centralizing patient data, making it attractive for ophthalmic imaging where privacy is paramount. However, annotation reliability at participating institutions cannot always be assumed. In ophthalmic imaging, differences in disease prevalence and class composition can resemble changes caused by corrupted supervision, creating a fundamental challenge: distinguishing genuine heterogeneity from annotation noise. This paper introduces OCT-FedSIR, a reliability-aware spectral framework for federated OCT classification under client-dependent annotation noise and heterogeneous data distributions. The framework addresses the critical need for trustworthy federated learning in medical imaging, where noisy labels can severely degrade model performance and compromise clinical decision-making.
OCT-FedSIR combines five key components: (1) class-balanced spectral estimation, (2) Stage-I logit adjustment, (3) complementary spectral descriptors, (4) selective spectral relabeling, and (5) noise-aware federated optimization. The framework operates on the principle that annotation noise manifests as spectral deviations in the logit space, which can be detected and corrected without discarding client data. The spectral estimation process can be formalized as follows: for each client , the class-balanced spectral descriptor is computed from the logits as
, where
denotes the spectral transform. The Stage-I logit adjustment applies a correction
, where is derived from the spectral descriptors. Selective spectral relabeling then identifies corrupted annotations by comparing spectral signatures across clients, using a threshold to flag potential noise. The noise-aware federated optimization aggregates client updates while downweighting those identified as noisy, ensuring robust model convergence. The overall architecture is illustrated in the following Mermaid diagram:
Why it matters
The findings highlight the importance of reliability-aware federated learning in medical imaging. OCT-FedSIR's ability to distinguish between genuine data heterogeneity and annotation noise is crucial for trustworthy collaborative learning. The spectral approach provides a principled way to detect and correct noisy labels, outperforming existing methods like RoFL and FedCorr. The high correction precision (91.3%) and low false-correction rate (3.5%) indicate that the framework can reliably recover corrupted annotations. Moreover, the superiority of retaining corrected clients over pruning them (by 9.30 percentage points) underscores the value of preserving data even when it is initially noisy. This has significant implications for real-world federated deployments, where discarding data from entire institutions can be wasteful and ethically problematic. Future work could extend OCT-FedSIR to other imaging modalities and explore adaptive thresholding for the spectral relabeling step. The framework's robustness under asymmetric noise, where the original FedSIR was less effective, further demonstrates its practical utility in diverse and unpredictable clinical environments.
Who should read this
CS practitioners and researchers
Opening member contentโฆ