DCPHA

DCPHA performs cross-modal retrieval for neuroimaging by embedding heterogeneous high-dimensional feature descriptors into a low-dimensional, consistency-preserving Hamming space.

Key Features:

• Multi-Manifold Property Utilization: Leverages the multi-manifold property of feature distributions to maintain continuous similarity across samples on manifolds.
• Asymmetric Auto-encoders and Attention Branches: Uses a pair of asymmetric auto-encoders with semantics-preserving attention branches during encoding and decoding to preserve semantic integrity.
• Multiple Pseudo-Siamese Network Encoder: Employs a multiple pseudo-Siamese network encoder to extract modality-specific features when more than two medical image modalities are present.
• Riemann Manifold-Based Similarity Definition: Defines continuous similarity for heterogeneous and homogeneous samples via multiple sub-manifolds on Riemann manifolds.
• Multi-Semantic Consistency and Multi-Manifold Similarity-Preserving Constraints: Learns hash codes under multi-semantic consistency and multi-manifold similarity-preserving constraints to enhance discriminability among semantically similar samples.

Scientific Applications:

• Cross-modal neuroimage retrieval: Enables retrieval across heterogeneous neuroimaging modalities, addressing modality gaps and semantic ambiguity in medical imaging datasets.

Methodology:

Embeds feature descriptors into a Hamming space using asymmetric auto-encoders with semantics-preserving attention branches; applies multiple pseudo-Siamese network encoders for multi-modality feature extraction; defines continuous similarity via multiple sub-manifolds on Riemann manifolds; and optimizes hash codes under multi-semantic consistency and multi-manifold similarity-preserving constraints.