Prototype-Decomposed Feature Imputation for Incomplete Multi-Field-of-View and Multi-Projection OCTA Image Classification.

Abstract

Optical coherence tomography angiography (OCTA) provides high-resolution visualization of fundus microvasculature and plays a crucial role in the diagnosis of fundus diseases. However, in real-world clinical practice, variations in imaging devices and acquisition protocols across centers lead to pronounced heterogeneity in OCTA data across field-of-view (FOV) and projection-layer dimensions, manifested as inconsistent FOV coverage and unavailable FOV-projection inputs. Existing methods predominantly focus on small-FOV images (e.g., $3\times 3$ and $6\times \text{6}~\text{mm}^{2}$), with limited utilization of larger-FOV information, making them less capable of handling data incompleteness and heterogeneity in complex clinical scenarios and thereby limiting their generalization and robustness. To address these issues, this paper presents a prototype-decomposed feature imputation framework for disease classification from incomplete multi-FOV and multi-projection OCTA data. By leveraging prototype-guided feature decomposition and imputation, the proposed framework enables disease classification through the dynamic integration of complementary representations. The framework comprises two stages: feature imputation and feature classification. In the feature-imputation stage, a prototype-based reconstruction network jointly employs contrastive learning and clustering to construct FOV-projection-specific prototypes. This design promotes a compact and discriminative encoded feature space while mitigating representation discrepancies arising from heterogeneous FOVs, projection layers, and imaging devices. Each available encoded feature is subsequently decomposed into a prototype-related shared component and a disease-related residual component, which are used to estimate the representation of each missing FOV-projection input through a non-parametric neighbor-based procedure. In the feature-classification stage, attention-based fusion mechanisms jointly model observed and imputed FOV-projection features across FOVs and projection layers. Furthermore, Dirichlet-based uncertainty modeling is incorporated to quantify the predictive uncertainty of the fused branches, followed by dynamic decision fusion to obtain the final classification result. The proposed framework is evaluated on a retrospective multi-center and multi-device OCTA cohort comprising 5,644 eyes. Comparative experiments with state-of-the-art feature-imputation and incomplete multi-input classification methods demonstrate that the proposed approach effectively estimates the representations of missing FOV-projection inputs and achieves superior performance in fundus disease classification.