Research map / Label-efficient and robust learning / Self-supervised and contrastive

Deep Hierarchies and Invariant Disease-Indicative Feature Learning for Computer Aided Diagnosis of Multiple Fundus Diseases

AAAI 2025 · Label-efficient and robust learning › Self-supervised and contrastive · approach: Self-supervised neuroimaging representations · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes a hierarchical vision transformer (HVT) with forward-attention over multi-level VGG19 features plus a self-supervised mask-consistent feature learner (MCFL, EMA teacher-student) for multi-label classification of six fundus conditions on ODIR.

Problem

Automated recognition of multiple co-occurring fundus diseases (DR, glaucoma, cataract, myopia retinopathy, AMD, normal) where lesion types overlap and vary across disease stages.

Approach

Combines a forward-attention module that propagates lower-level CNN features into higher-level tokens with a mean-teacher style masked feature-consistency objective, where an EMA global learner produces classification-guided pseudo-label features that a masked local learner must reproduce.

Key results

• On the ODIR off-site test set, the proposed method reaches F1 95.9, accuracy 93.1 and Kappa 75.5, versus 95.0/91.8/71.1 for the best CNN+ViT baseline.
• Ablation on the on-site set: the VGG19 MLFE baseline gets Kappa 64.1, adding MCFL gives 64.8, adding HVT gives 65.6, and both give 67.7.
• The 16x16 mask scale and 50% mask ratio gave the best reported accuracy of 93.13% in the hyperparameter analysis.

Measured effects reported in the paper

Limitations

• Evaluated only on a single dataset (ODIR) with six classes.
• Gains over baselines are small on the on-site test set (F1 94.8 vs 95.0 for VGG19 in the table).
• Segmentation of lesions is left for future work; no lesion-level evaluation.

Experimental setup

Datasets: ODIR

Benchmarks: ODIR off-site testing set ODIR on-site testing set

Metrics: F1 accuracy Kappa score

Baselines: ResNet50 EfficientNet_B7 VGG16 ViT-L/16 Swin-B BEiT PVT CvT-13 Hiera-MAE-tiny Hiera-MAE-base XCiT CrossViT

Builds on: masked autoencoder mean teacher CrossViT VGG-19

Closest papers by meaning

• MSSDA: Multi-Sub-Source Domain Adaptation for Diabetic Foot Neuropathy Recognition (AAAI 2025 · similarity 0.84)
• FOCUS: Knowledge-enhanced Adaptive Visual Compression for Few-shot Whole Slide Image Classification (CVPR 2025 · similarity 0.84)
• PathVQ: Reforming Computational Pathology Foundation Model for Whole Slide Image Analysis via Vector Quantization (NeurIPS 2025 · similarity 0.84)
• SIGraph: Saliency Image-Graph Network for Retinal Disease Classification in Fundus Image (AAAI 2025 · similarity 0.83)
• Like an Ophthalmologist: Dynamic Selection Driven Multi-View Learning for Diabetic Retinopathy Grading (AAAI 2025 · similarity 0.83)
• Decomposing Disease Descriptions for Enhanced Pathology Detection: A Multi-Aspect Vision-Language Pre-training Framework (CVPR 2024 · similarity 0.83)