CLEAR-Mamba:Towards Accurate, Adaptive and Trustworthy Multi-Sequence Ophthalmic Angiography Classification

TL;DR

This work tackles the challenge of reliable, generalizable classification in single-modality ophthalmic angiography by exploiting temporal dynamics and uncertainty-aware decision making. The authors introduce CLEAR-Mamba, an encoder-adaptor-predictor framework built on MedMamba, incorporating a HyperNetwork-based HyperAdaptive Conditioning (HaC) and a reliability-focused Evidential Prediction (RaP) head to produce calibrated probabilities and uncertainty estimates. A large-scale in-house dataset with 43 disease categories across FFA/ICGA sequences is constructed via a multi-agent data engine, enabling robust evaluation and ablation studies. Across in-house and public benchmarks, CLEAR-Mamba consistently outperforms CNN/ViT/Mamba baselines in accuracy and reliability, demonstrating improved temporal utilization, generalization, and confidence calibration for clinically practical angiography classification.

Abstract

Medical image classification is a core task in computer-aided diagnosis (CAD), playing a pivotal role in early disease detection, treatment planning, and patient prognosis assessment. In ophthalmic practice, fluorescein fundus angiography (FFA) and indocyanine green angiography (ICGA) provide hemodynamic and lesion-structural information that conventional fundus photography cannot capture. However, due to the single-modality nature, subtle lesion patterns, and significant inter-device variability, existing methods still face limitations in generalization and high-confidence prediction. To address these challenges, we propose CLEAR-Mamba, an enhanced framework built upon MedMamba with optimizations in both architecture and training strategy. Architecturally, we introduce HaC, a hypernetwork-based adaptive conditioning layer that dynamically generates parameters according to input feature distributions, thereby improving cross-domain adaptability. From a training perspective, we develop RaP, a reliability-aware prediction scheme built upon evidential uncertainty learning, which encourages the model to emphasize low-confidence samples and improves overall stability and reliability. We further construct a large-scale ophthalmic angiography dataset covering both FFA and ICGA modalities, comprising multiple retinal disease categories for model training and evaluation. Experimental results demonstrate that CLEAR-Mamba consistently outperforms multiple baseline models, including the original MedMamba, across various metrics-showing particular advantages in multi-disease classification and reliability-aware prediction. This study provides an effective solution that balances generalizability and reliability for modality-specific medical image classification tasks.

Figures (7)

• Figure 1: Automated pipeline for extracting, anonymizing, and aligning medical image–text data from raw PDF reports.
• Figure 2: Dataset statistics. (a) Class distribution across 43 ocular categories showing a long-tailed pattern. (b) Modality proportion between FFA and ICGA images.
• Figure 3: CLEAR-Mamba Framework.
• Figure 4: Component-level ablation study of the CLEAR framework. (a) Comparison of overall accuracy (OA) and AUC among different component configurations. (b) Boxplots and (c) density distributions of prediction confidence for correctly and incorrectly classified samples in different variants (HaC, RaP, and full model).
• Figure 5: Hyperparameter-level ablation study of the CLEAR framework. Subfigures (a–d) are arranged clockwise from the top-left. (a) Effect of the hypernetwork feature dimension (had_feat_dim); (b) impact of the compression ratio (r); (c) sensitivity to the adaptive evidence coefficient ($\lambda_e$); and (d) comparison of EDL update strategies.