Source-linked AI summary
RTNet: Relation Transformer Network for Diabetic Retinopathy Multi-lesion Segmentation
Shiqi Huang, Jianan Li, Yuze Xiao, Ning Shen, Tingfa Xu
TL;DR
DR lesion segmentation could assist diagnosis, but prior research gave limited attention to pathological associations among lesions and vessels. RTNet uses dual-branch GTB and RTB modules to preserve details and model lesion–vessel and lesion–lesion relationships. On IDRiD and DDR, it reports superior or competitive multi-lesion segmentation performance, while coarse vessel pseudo masks remain a limitation.
Problem
DR lesion segmentation is challenging, and prior approaches paid insufficient attention to pathological associations among lesions and vessels despite their relevance to diagnosis assistance.
Method
RTNet uses dual branches with GTB for detailed lesion features and RTB self- and cross-attention to model lesion–lesion and lesion–vessel relationships.
Results
The network segments four DR lesion types simultaneously and outperforms reported state-of-the-art methods on IDRiD and DDR, with best performance reported for EX, MA, and SE.
Takeaways & Limitations
Pathological relationships among lesions and vessels provide an exploitable basis for multi-lesion segmentation in fundus images.
Takeaways & Limitations
Coarse vessel pseudo masks from semi-supervised learning limit the network because pixel-level vessel annotations are costly.
Abstract
from arXiv · showhide
Automatic diabetic retinopathy (DR) lesions segmentation makes great sense of assisting ophthalmologists in diagnosis. Although many researches have been conducted on this task, most prior works paid too much attention to the designs of networks instead of considering the pathological association for lesions. Through investigating the pathogenic causes of DR lesions in advance, we found that certain lesions are closed to specific vessels and present relative patterns to each other. Motivated by the observation, we propose a relation transformer block (RTB) to incorporate attention mechanisms at two main levels: a self-attention transformer exploits global dependencies among lesion features, while a cross-attention transformer allows interactions between lesion and vessel features by integrating valuable vascular information to alleviate ambiguity in lesion detection caused by complex fundus structures. In addition, to capture the small lesion patterns first, we propose a global transformer block (GTB) which preserves detailed information in deep network. By integrating the above blocks of dual-branches, our network segments the four kinds of lesions simultaneously. Comprehensive experiments on IDRiD and DDR datasets well demonstrate the superiority of our approach, which achieves competitive performance compared to state-of-the-arts.
I. INTRODUCTION
DR lesion segmentation can assist diagnosis amid limited ophthalmologist availability, but prior work largely emphasized network design rather than pathological lesion connections. RTNet addresses this gap with dual-branch transformer blocks that use lesion, vessel, and global feature relationships, achieving strong multi-lesion results on IDRiD and DDR.
- Motivation: DR lesions support diagnosis, while screening burdens primary-care systems because ophthalmologists are in short supply.Automatic lesion segmentation is presented as assistance for ophthalmologists.
- Research gap: Pathological connections between lesions and vessels have received insufficient attention in prior complex medical segmentation research.The paper highlights both lesion–vessel associations and spatial interactions among lesions.
- Proposed method: RTB combines cross-attention between lesion and vessel features with self-attention over lesion features to exploit pathological dependencies.The dual-branch design uses vascular information to help locate lesions and lesion relationships to address similar-looking classes.
- Proposed method: GTB preserves detailed information and generates separate vessel and lesion features for small lesions and blurred borders.The two GTB outputs are subsequently investigated at the pathological-connection level by RTB.
- Results: The network segments four lesion types simultaneously and outperforms reported state-of-the-art methods on IDRiD and DDR, with best performance reported for EX, MA, and SE.Ablation experiments on IDRiD also evaluate the contributions of RTB and GTB.
II. RELATED WORK
DR lesion segmentation is difficult because lesions vary morphologically across disease stages, are often tiny, and resemble other fundus structures. The paper motivates exploiting pathological spatial associations while noting that existing approaches either overlook these connections or separate tasks at substantial computational cost.
- Challenges: Large intra-class variation and disease-stage differences make DR lesion segmentation challenging.The task also involves pathological associations that appear in lesion spatial distributions.
- Pathological associations: EX lesions often surround MAs, SE lesions commonly border HEs, and vascular abnormalities are linked to lesion distributions.The paper presents these patterns as pathological relationships observed in fundus images.
- Pathological associations: Distance statistics on IDRiD verify exploitable distribution patterns among EX, MA, SE, HE, capillaries, and arteries.The analysis measures distances between lesion clusters and nearby lesions or vascular structures.
- Existing limitations: Traditional methods face morphological variation and lesion–structure confusion, while deep networks can lose details from lesions measuring only one or two pixels.These issues motivate attention mechanisms and feature-fusion strategies.
- Related approaches: Prior directions either fuse attention features to preserve details or separate red-lesion and exudate segmentation to reduce inter-class imbalance.The latter strategy can require time and substantial memory, according to the paper.
C. Transformer in Medical Images
Transformers moved from NLP into computer vision and medical image processing, where they support global or multi-scale feature modeling for segmentation and prediction tasks.
- Transformer networks use self-attention for efficient parallel computation and have supported multiple NLP tasks since 2017.They were introduced as an alternative to recurrent architectures in the cited background.
- Vision Transformer adapts the transformer to images by treating linearly encoded image patches as tokens.
- Medical imaging studies have applied transformer variants to microscopy prediction and anisotropic 3D segmentation.The cited examples include global pixel transformers and ViT-based architectures within U-Net.
- The proposed work is organized around two key components, the Global Transformer Block and Relation Transformer Block, followed by a designed loss function.
A. Overview
The network uses a dual-branch architecture to learn vessel and lesion features separately, combining global context, detailed channel-spatial refinement, and pathological relation modeling for parallel segmentation.
- Overview: The network outputs one vascular mask and four lesion masks through backbone, GTB, RTB, and segmentation-head components.
- Overview: GTB branches transform shared backbone features into vessel and lesion representations with long-range contextual information.
- Global Transformer Block: GTB uses query, key, and value generators to produce attentively refined feature maps for lesions and vessels.The query uses convolution and global average pooling, while key and value use reshaped representations.
- Global Transformer Block: GTB fuses channel-wise and spatial-wise weighted features with the input through a residual connection, preserving the input format while enriching object-specific features.
- Global Transformer Block: GTB targets small discrete lesion patterns by highlighting useful information concentrated in only a few pixels of selected channels.
- Overview: The pipeline applies GTB to both branches and RTB afterward to model relations among lesions and between lesions and vessels.
C. Relation Transformer Block
RTB combines self-attention over lesion features with cross-attention from lesion queries to vessel features, modeling lesion dependencies and vascular interactions for segmentation refinement.
- Relation Transformer Block: RTB contains self-attention and cross-attention heads for intra-class lesion dependencies and inter-class lesion-vessel relations.
- Relation Transformer Block: The self-attention head derives query, key, and value from lesion features, whereas cross-attention derives its key from vessel features.
- Relation Transformer Block: The two attentive outputs use residual learning and are concatenated along the channel dimension to form the RTB output.
- Self-attention: Self-attention exchanges information across all positions, modeling long-range relations among dispersed lesions and refining mixtures and large-pattern edges.
- Cross-attention: Cross-attention queries global vascular structures to better locate MA and SE while reducing false positives involving EX reflections and MA-capillary confusion.
D. Loss Function
The network jointly optimizes multi-lesion and vessel segmentation using a weighted sum of their respective losses, with λ controlling vascular supervision.
- The total loss is the sum of lesion and vessel losses, weighted by λ for the vessel branch.
- The lesion loss is a five-class weighted cross-entropy, while the vessel loss is a binary weighted cross-entropy.
- When λ is zero, optimization uses multi-lesion features only; increasing λ increases the role of vascular information.
IV. EXPERIMENTS AND RESULTS
The study evaluates RTNet on the IDRiD and DDR datasets, which provide pixel-level annotations for four DR lesion types.
- IDRiD contains 81 fundus images with pixel-level annotations for EX, HE, MA, and SE lesions.
- DDR provides 757 fundus images with pixel-level annotations for EX, HE, MA, and SE lesions.
B. Implementation Details
The implementation combines vessel-aware preprocessing, augmentation, a DenseNet-161 UNet backbone, and PR/ROC AUC evaluation with comparisons on IDRiD and DDR.
- Training data are augmented through vessel-mask generation, 512 × 512 random crops, flips, rotations, and CLAHE contrast enhancement.
- DenseNet-161 pretrained on ImageNet is used as the UNet encoder with output channel number C set to 32.
- Models are trained for 250 epochs with SGD, batch size 16, initial learning rate 0.001, momentum 0.9, and weight decay 0.0005.
- Performance is evaluated using AUC for precision-recall and receiver operating characteristic curves, with greater emphasis on recall for lesion segmentation.
- Comparisons on IDRiD: On IDRiD, the method ranks first for AUC ROC of EX, MA, and SE and AUC PR of EX and SE, while ranking second for HE on both metrics.
- Comparisons on DDR: On DDR, the method achieves the best performance for EX, MA, and SE and ranks second for HE.
E. Ablation Studies on IDRiD Dataset
Ablation studies examine backbone choice, vessel regularization, GTB, attention components, and their combined effect on four-lesion segmentation.
- Backbone selection: DenseNet-161 integrated with UNet achieves the best performance across all lesions among the evaluated backbone architectures.
- Regularization terms: Rthin improves MA segmentation, while Rcl achieves the best AUC PR results for HE and MA.
- Regularization terms: Vessel regularization produces limited final improvement because vessel ground truths are semi-supervised pseudo-masks rather than manual annotations.
- GTB analysis: GTB improves segmentation for all four lesions and outperforms other attention blocks under the same model parameters.
- GTB analysis: GTB highlights discrete and small patterns that CBAM spatial attention overlooks.
3) Analyze the Effect of GTB:
RTB uses lesion self-attention and vessel cross-attention to focus on complementary structures, and combining RTB with GTB produces the strongest ablation result.
- RTB attention: RTB visualizations show lesion-specific query positions and distinct self-attention and cross-attention maps for EX, HE, MA, and SE.
- RTB attention: Cross-attention specializes in vascular tributaries, while self-attention focuses on smaller patterns for MA queries.
- RTB attention: Both self-attention and cross-attention contribute to improved network performance.
- Combined model: The complete network combining GTB and RTB achieves the highest results, with its Fig. 5 curve nearly enclosing the alternatives.
F. Generalization Studies on DDR and IDRiD Dataset
The study evaluates cross-dataset generalization from DDR to IDRiD and qualitatively examines how GTB and RTB affect segmentation. The reported visual results associate GTB with recovering missing detections and RTB with further improvements, fewer false alarms, and sharper lesion boundaries.
- Generalization study: The generalization study trains models on DDR images and tests them on IDRiD images captured from another source.This setup is used to assess generalization across imaging domains.
- Generalization study: The authors report that their method achieves the best performance in the cross-dataset generalization comparison.
- Qualitative analysis: GTB recovers missing lesion detections relative to the baseline in qualitative segmentation results.The visualization marks these recoveries with yellow boxes.
- Qualitative analysis: RTB further recovers missing detections and reduces false alarms relative to baseline+GTB.The visualization marks these changes with green boxes.
- Qualitative analysis: RTB more precisely fine-tunes the edges of large lesions, especially soft exudates with blurred edges.
- Limitation: The authors identify coarse vessel pseudo masks from semi-supervised learning as a limitation that makes the network inadequate.They link this limitation to the high cost of expert pixel-level annotations.