AO-SCA: Automated Open/Narrow/Synechiae ACA

• AO-SCA is an automated system for classifying anterior chamber angle states (open, narrow, synechiae) in AS-OCT scans using sequence modeling and multi-scale features.
• It leverages a customized SMA-Net architecture with an Xception backbone, MSDA blocks, SE modules, and ConvLSTM layers to capture both spatial and temporal dependencies.
• The method demonstrates state-of-the-art performance with significant gains in kappa, F1 score, and AUC compared to conventional 2D and 3D CNN baselines.

The abbreviation AO-SCA appears in distinct contexts across research domains, notably in automated medical image analysis for anterior chamber angle (ACA) classification in ophthalmology, as well as in adaptive optics and communications. In the setting of anterior segment optical coherence tomography (AS-OCT), AO-SCA specifically refers to the automated classification of ACA states—Open, Narrow, or Synechiae—using sequence modeling and multi-scale features. This article focuses on AO-SCA as detailed in "Open-Narrow-Synechiae Anterior Chamber Angle Classification in AS-OCT Sequences" (Hao et al., 2020).

1. Problem Definition and Clinical Significance

The anterior chamber angle (ACA) is a critical anatomical structure for the regulation of intraocular pressure; its configuration underpins the diagnosis and management of angle-closure glaucoma. Traditional automated procedures have approached ACA classification as a binary open/closed problem. However, a three-class system (open, narrow, synechiae) is essential clinically, as it allows differentiation between pre-synechial (appositional) and true synechial closure—crucial for prognosis and intervention strategies. AO-SCA, in this context, stands for automated open-narrow-synechiae ACA classification in AS-OCT, aiming to construct an accurate, reproducible, and clinically informative three-way grading from sequential image data (Hao et al., 2020).

2. Data Acquisition and Preprocessing

AO-SCA workflows hinge on sequence data derived from volumetric AS-OCT scans:

• An AS-OCT sequence consists of $T=11$ adjacent B-scans spanning a $15^\circ$ gonioscopic sweep. Each original slice is $2144\times1876$ pixels.
• The ACA region is localized via a coarse-to-fine strategy, initially cropping to $448\times448$ pixels and then center-cropping each slice to $224\times224$.
• Dataset specifics: 66 eyes imaged on a CASIA-2 system, yielding 1584 annotated ACA segments. Labels are "open," "narrow," or "synechiae" (class counts: 504, 742, 338) (Hao et al., 2020).

This tailored preprocessing ensures that only the anatomical region of interest is fed to downstream classification networks, maximizing both computational efficiency and clinical relevance.

3. Network Architecture: SMA-Net for AO-SCA

The proposed state-of-the-art network for AO-SCA is SMA-Net, which incorporates both spatial and temporal modeling:

• Backbone: Xception, a deep CNN with depthwise-separable convolutions, is employed for slice-level feature extraction.
• Multi-Scale Discriminative Aggregation (MSDA) Block: All $3\times3$ convolutions are replaced with MSDA modules, comprising three parallel atrous (dilated) convolutions with dilation rates $\{1,2,3\}$ in a hierarchical residual configuration.
• Squeeze-and-Excitation (SE) gate: Fused multi-scale features undergo global average pooling followed by a channel-wise scaling (via a two-layer MLP with ReLU and sigmoid activations).
• ConvLSTM: Two stacked ConvLSTM layers ($1024$ hidden channels each) operate sequentially across the $T$ input slices, preserving spatiotemporal dependencies.
• Dual prediction heads: Each ConvLSTM state yields both a slice-level (per $Y_t$ via softmax) and a sequence-level (aggregated $15^\circ$0) prediction (Hao et al., 2020).

MSDA and SE modules enable robust spatial discrimination, while ConvLSTM captures anatomical evolution across the gonioscopic sequence—a requirement for distinguishing subtle morphologic transitions, especially between narrow and synechial states.

4. Training Protocol, Loss Function, and Optimization

The SMA-Net employs a composite loss to jointly optimize slice-level and sequence-level predictions:

$15^\circ$1

with $15^\circ$2.

• $15^\circ$3: summed cross-entropy over all slices, encouraging intra-sequence consistency.
• $15^\circ$4: cross-entropy on the final sequence-aggregated prediction.
• Training uses the Adam optimizer, an initial learning rate of $15^\circ$5 with decay on plateau, and per-epoch brightness, contrast, color, and sharpness augmentation (Hao et al., 2020).

Multi-level supervision is key: it stabilizes optimization and enables the network to capture both local and global features, a necessity for clinically valid AO-SCA.

5. Quantitative Performance and Ablation

AO-SCA, as realized through SMA-Net, demonstrates statistically significant superiority over both 2D and 3D CNN baselines:

Open-Narrow-Synechiae (Private Dataset, Test Split):

Method	$15^\circ$6	F1	B-Acc	Sens	Spec
ResNet-34 (2D)	0.6766	0.7527	0.8188	0.7485	0.8891
Xception (2D)	0.7252	0.7835	0.8393	0.7752	0.9035
MA-Net (MSDA only)	0.7477	0.8121	0.8600	0.8063	0.9137
C3D (3D)	0.7489	0.8115	0.8532	0.8048	0.9136
I3D (3D)	0.7662	0.8171	0.8619	0.8073	0.9166
SMA-Net	0.7931	0.8459	0.8829	0.8371	0.9282

AUC for narrow vs. synechiae: SMA-Net achieves 0.8207, outperforming all tested alternatives (Hao et al., 2020).

Ablation demonstrates additive performance gains from MSDA, ConvLSTM, and the multi-loss structure—each contributing 2–4% to $15^\circ$7 and AUC.

6. Core Methodological Innovations

The critical elements distinguishing AO-SCA via SMA-Net are:

• MSDA Block: Hierarchical multi-scale feature extraction via parallel atrous convolutions and residual summation captures fine-to-coarse anatomical cues, essential in ophthalmic imaging.
• Temporal Modeling: ConvLSTM explicitly encodes anatomical progression, critical for distinguishing between stages of angle closure not discernible from single slices.
• Dual Loss: Synchronous optimization for per-slice and sequence-level classifications ensures both local and global correctness, mitigating issues such as vanishing gradients.

These components, integrated via a fully differentiable architecture, collectively enable robust, end-to-end three-way AO-SCA classification (Hao et al., 2020).

7. Clinical and Technical Implications

AO-SCA provides clinically more relevant information than binary open/closed classifiers, potentially informing early intervention strategies in primary angle closure suspect (PACS) versus synechial disease. The fusion of multi-scale spatial encoding and temporal modeling demonstrated by SMA-Net is thus not only technically optimal (state-of-the-art on available datasets) but also aligns with ophthalmological grading standards.

A plausible implication is that as sequence-based approaches become standard, the AO-SCA paradigm can improve both screening throughput and diagnostic accuracy in AS-OCT-based glaucoma care, with performance verified on both private and public benchmarks (top scores in all metrics on the AGE dataset as well) (Hao et al., 2020).