Research map / Detection and segmentation / CNN features
HACDR-Net: Heterogeneous-Aware Convolutional Network for Diabetic Retinopathy Multi-Lesion Segmentation
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 HACDR-Net, a convolutional encoder-decoder with heterogeneous cross-convolution and modulated deformable convolution branches, plus NALoss (Gaussian-noise-perturbed loss with hard example mining) for diabetic retinopathy multi-lesion segmentation on IDRiD and DDR.
Problem
Diabetic retinopathy multi-lesion segmentation is hard because lesions vary widely in shape and size and occupy a very small fraction of fundus images, creating severe foreground-background imbalance.
Approach
Aggregates multi-branch heterogeneous cross-convolutions and deformable convolutions via an attention-based aggregation module (HAAA) to get heterogeneous receptive fields, and balances training with NALoss, which adds class-weighted Gaussian noise to predicted logits and then applies hard example mining.
Key results
- On DDR, HACDR-Net improves over the previous best M2MRF by 3.53% mDice, 3.29% mIoU and 0.94% mAUPR.
- On IDRiD, HACDR-Net improves mAUPR by 2.79%, mDice by 0.26% and mIoU by 0.56% over M2MRF.
- Ablation on DDR: removing HAAA drops mAUPR by 4.66%; NALoss beats weighted cross-entropy and OHEM (mDice 49.30 vs 47.88 and 47.12).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace M2MRF with HACDR-Net | DR multi-lesion segmentation · IDRiD | mAUPR: 66.00 -> 68.79 | Helped |
| replace M2MRF with HACDR-Net | DR multi-lesion segmentation · DDR | mDice: 45.77 -> 49.30 | Helped |
| add the HAAA module (ablation) | HACDR-Net · DDR | mDice: 45.31 -> 49.30 | Helped |
| add the DFFN module (ablation) | HACDR-Net · DDR | mDice: 49.09 -> 49.30 | Helped |
| add the ONFA branch to MLKCC, DCC, and MDConv (ablation) | HACDR-Net · DDR | mDice: 49.04 -> 49.30 | Helped |
| add all heterogeneous branches with MDConv rather than using MDConv alone (ablation) | HACDR-Net · DDR | mDice: 46.55 -> 49.30 | Helped |
| include 1×1 convolution, 5×5 MDConv, and attention together in HAAA (ablation) | HACDR-Net · DDR | mDice: 49.10 -> 49.30 | Helped |
| replace NALoss with Cross-Entropy loss (ablation) | HACDR-Net · DDR | mDice: 49.30 -> 47.04 | Hurt |
| replace NALoss with Weighted Cross-Entropy loss (ablation) | HACDR-Net · DDR | mDice: 49.30 -> 47.88 | Hurt |
| replace NALoss with Dice loss (ablation) | HACDR-Net · DDR | mDice: 49.30 -> 46.50 | Hurt |
| replace NALoss with OHEM loss (ablation) | HACDR-Net · DDR | mDice: 49.30 -> 47.12 | Hurt |
| combine NALoss with Weighted Cross-Entropy instead of using NALoss alone (ablation) | HACDR-Net · DDR | mDice: 49.30 -> 49.08 | Hurt |
Surprising findings
- On IDRiD, HACDR-Net is slightly below M2MRF on IoU for some categories and its mDice gain over M2MRF is only 0.26%.
Limitations
- Gains over the best prior method on IDRiD are small (e.g., 0.26% mDice).
- Evaluated only on two datasets with four lesion types.
- NALoss requires a two-stage training schedule and tuned threshold and noise parameters.
Experimental setup
Datasets: IDRiD DDR
Benchmarks: IDRiD DDR
Metrics: IoU mIoU Dice mDice AUPR mAUPR
Baselines: HED PSPNet DenseUNet Deeplabv3+ L-seg DNL HRNetV2 Twins-SVT-B TransUnet Swin-Unet Swin-Tv2 PMCNet
Builds on: Visual Attention Network (VAN) U-Net modulated deformable convolution online hard example mining M2MRF
Closest papers by meaning
- Like an Ophthalmologist: Dynamic Selection Driven Multi-View Learning for Diabetic Retinopathy Grading (AAAI 2025 · similarity 0.86)
- SIGraph: Saliency Image-Graph Network for Retinal Disease Classification in Fundus Image (AAAI 2025 · similarity 0.86)
- FOCUS: Benchmarking Retinal Model Generalization from Foundation Vision Encoders to Multimodal LLMs (NeurIPS 2026 · similarity 0.85)
- Domain Adaptive Diabetic Retinopathy Grading with Model Absence and Flowing Data (CVPR 2025 · similarity 0.83)
- DDViT: Double-Level Fusion Domain Adapter Vision Transformer (Student Abstract) (AAAI 2024 · similarity 0.83)
- Minding Fuzzy Regions: A Data-driven Alternating Learning Paradigm for Stable Lesion Segmentation (CVPR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.