[Paper Review] BSDA-Net: A Boundary Shape and Distance Aware Joint Learning Framework for Segmenting and Classifying OCTA Images

Li Lin, Zhonghua Wang+6 more|arXiv (Cornell University)|Jul 10, 2021

Retinal Imaging and Analysis30 references4 citations

TL;DR

This paper proposes BSDA-Net, a novel joint learning framework that enhances FAZ segmentation and diagnostic classification in OCTA images by integrating boundary heatmap regression and signed distance map reconstruction branches. The method achieves state-of-the-art performance on three datasets, with 94.23% accuracy and 87.68% Kappa on OCTA-500, demonstrating the effectiveness of FAZ features in boosting classification through hierarchical feature fusion.

ABSTRACT

Optical coherence tomography angiography (OCTA) is a novel non-invasive imaging technique that allows visualizations of vasculature and foveal avascular zone (FAZ) across retinal layers. Clinical researches suggest that the morphology and contour irregularity of FAZ are important biomarkers of various ocular pathologies. Therefore, precise segmentation of FAZ has great clinical interest. Also, there is no existing research reporting that FAZ features can improve the performance of deep diagnostic classification networks. In this paper, we propose a novel multi-level boundary shape and distance aware joint learning framework, named BSDA-Net, for FAZ segmentation and diagnostic classification from OCTA images. Two auxiliary branches, namely boundary heatmap regression and signed distance map reconstruction branches, are constructed in addition to the segmentation branch to improve the segmentation performance, resulting in more accurate FAZ contours and fewer outliers. Moreover, both low-level and high-level features from the aforementioned three branches, including shape, size, boundary, and signed directional distance map of FAZ, are fused hierarchically with features from the diagnostic classifier. Through extensive experiments, the proposed BSDA-Net is found to yield state-of-the-art segmentation and classification results on the OCTA-500, OCTAGON, and FAZID datasets.

Motivation & Objective

• To address the challenge of imprecise FAZ boundary detection and outlier generation in OCTA image segmentation due to limited prior knowledge in single-task learning.
• To explore whether explicit FAZ-related features (shape, boundary, signed distance) can improve deep diagnostic classification performance in OCTA images.
• To develop a joint learning framework that hierarchically fuses segmentation and classification features for improved clinical diagnostic accuracy.
• To establish a new SOTA baseline for both FAZ segmentation and multi-disease classification using OCTA images.

Proposed method

• Introduces two auxiliary branches—boundary heatmap regression and signed distance map (SDM) reconstruction—alongside the main segmentation branch to enhance boundary precision and reduce outliers.
• Uses a shared encoder followed by three parallel branches: segmentation, boundary heatmap regression, and SDM reconstruction, enabling multi-level feature learning.
• Applies a soft contour and directional signed distance-aware loss function derived from the auxiliary branches to guide segmentation with improved boundary awareness.
• Employs hierarchical feature fusion by combining low-level and high-level features from all three branches with features from the diagnostic classifier, enabling interpretable feature integration.
• Utilizes ResNeSt50 as the backbone network and applies joint training with multi-task loss to optimize segmentation and classification simultaneously.
• Applies a multi-level perception strategy that fuses shape, boundary, and signed directional distance features at multiple scales to enrich the classifier's input representation.

Experimental results

Research questions

• RQ1 Can auxiliary branches for boundary heatmap regression and signed distance map reconstruction significantly improve FAZ segmentation accuracy and boundary precision in OCTA images?
• RQ2 To what extent do explicit FAZ features (shape, boundary, signed distance) enhance the performance of deep diagnostic classification networks for ocular diseases?
• RQ3 Does a hierarchical joint learning framework that fuses segmentation and classification features outperform single-task or standard multi-task learning in OCTA image analysis?
• RQ4 How does the proposed BSDA-Net compare to existing SOTA methods in terms of segmentation and classification performance across diverse OCTA datasets?

Key findings

• BSDA-Net achieves 94.23% accuracy and 87.68% Kappa on the OCTA-500 dataset, outperforming all compared methods including ResNeSt50 and YNet.
• On the OCTAGON dataset, BSDA-Net achieves 99.53% accuracy and 98.92% Kappa, setting a new SOTA benchmark for classification in OCTA imaging.
• On the FAZID dataset, BSDA-Net achieves 82.57% accuracy and 73.67% Kappa, with particularly strong performance on diabetic retinopathy (89.32% F1-score).
• The ablation study confirms that removing the boundary (B) and distance (D) branches reduces classification accuracy by 4.5 percentage points, proving their critical role.
• The model shows robustness in detecting FAZ morphology changes, with high recall for diabetic and AMD cases, indicating clinical relevance.
• Visualization results confirm that the predicted soft boundaries and reconstructed SDMs are more accurate and less noisy than baseline methods, especially in low-contrast regions.