[Paper Review] Connection Sensitive Attention U-NET for Accurate Retinal Vessel Segmentation

Ruirui Li, Mingming Li+2 more|arXiv (Cornell University)|Mar 13, 2019

Retinal Imaging and AnalysisMedicine25 references36 citations

TL;DR

Introduces CSAU, a retinal vessel segmentation model that combines a connection sensitive loss with attention gates to preserve thin vessels and boundaries, achieving leading results on DRIVE, STARE, and HRF.

ABSTRACT

We develop a connection sensitive attention U-Net(CSAU) for accurate retinal vessel segmentation. This method improves the recent attention U-Net for semantic segmentation with four key improvements: (1) connection sensitive loss that models the structure properties to improve the accuracy of pixel-wise segmentation; (2) attention gate with novel neural network structure and concatenating DOWN-Link to effectively learn better attention weights on fine vessels; (3) integration of connection sensitive loss and attention gate to further improve the accuracy on detailed vessels by additionally concatenating attention weights to features before output; (4) metrics of connection sensitive accuracy to reflect the segmentation performance on boundaries and thin vessels. Our method can effectively improve state-of-the-art vessel segmentation methods that suffer from difficulties in presence of abnormalities, bifurcation and microvascular. This connection sensitive loss tightly integrates with the proposed attention U-Net to accurately (i) segment retinal vessels, and (ii) reserve the connectivity of thin vessels by modeling the structural properties. Our method achieves the leading position on DRIVE, STARE and HRF datasets among the state-of-the-art methods.

Motivation & Objective

• Motivate accurate segmentation of retinal vessels with emphasis on thin and boundary regions.
• Develop a loss function that encodes connectivity and local structural properties to guide pixel-wise segmentation.
• Integrate an attention mechanism to enhance learning of fine vascular structures at skip connections.
• Propose a new evaluation metric, connection sensitive accuracy, to reflect boundary and thin-vessel performance.
• Demonstrate state-of-the-art performance on public retinal vessel benchmarks (DRIVE, STARE, HRF).

Proposed method

• Propose a four-block encoder-decoder U-Net similar to Attention U-Net with enhanced skip connections.
• Introduce the connection sensitive loss L_cs that weights pixel-wise cross-entropy by connectivity-based coefficients θ1 and θ2 derived from local density C_i.
• Define C_i via a local 5x5 density estimation and Monte Carlo sampling to model connectivity probability in patches, guiding θ1, θ2 in L_cs.
• Incorporate an attention gate with UP-Link architecture to compute refined gating coefficients and concatenate final attention weights to features before output.
• Train end-to-end with AdamW, using data augmentation, multi-dataset validation, and a dynamic learning-rate strategy.
• Introduce a new accuracy metric ACC_cs that emphasizes boundary and thin-vessel correctness using a connectivity-aware mask.

Experimental results

Research questions

• RQ1 Can a connectivity-aware loss improve preservation of thin retinal vessels and boundary details beyond standard pixel-wise losses?
• RQ2 Does integrating a novel attention gate with connectivity guidance yield better vessel delineation, especially at bifurcations and crossings?
• RQ3 How does CSAU perform on established retinal vessel benchmarks (DRIVE, STARE, HRF) compared with state-of-the-art methods?
• RQ4 Is the proposed connection sensitive accuracy metric effective for evaluating boundary and thin-vessel segmentation performance?
• RQ5 What is the relative contribution of the CS loss and the attention gate to overall performance?

Key findings

• CSAU achieves leading F1-score, ROC AUC, and sensitivity across DRIVE, STARE, and HRF datasets among reported methods.
• On DRIVE, CSAU improves F1-score by up to 0.? over VGAN and achieves higher ACC_cs than DRIU, indicating better boundary and thin-vessel segmentation.
• On STARE, CSAU attains the best metrics with notable gains in ROC AUC and ACC_cs compared to baselines.
• On HRF, CSAU delivers the highest ROC AUC and competitive PR AUC and F1-score relative to published methods.
• A new connection sensitive accuracy metric ACC_cs reflects improvements in boundaries and thin vessels, aligning with qualitative benefits observed in segmentation outputs.
• Ablation studies show that both the CS loss and the attention mechanism contribute to performance gains, with the combination yielding the best results.