TL;DR
CLEAR-Mamba is a new deep learning framework designed to classify ophthalmic angiography (eye blood flow imaging) with high accuracy and reliable uncertainty estimation. By combining the efficiency of Mamba-based architectures with dynamic parameter adaptation and evidential learning, it significantly outperforms standard models on a massive new 43-disease dataset and several public benchmarks.
What Problem Does This Paper Address?
Ophthalmic angiography, specifically Fluorescein Fundus Angiography (FFA) and Indocyanine Green Angiography (ICGA), provides a "movie" of blood flow in the eye. Unlike static fundus photos, these sequences reveal how lesions evolve over time. However, automating their analysis is difficult for three main reasons:
- Temporal Neglect: Most existing AI models treat angiography as static images, missing critical hemodynamic signatures (like early filling or late leakage).
- Domain Shift: Significant variability between different imaging devices makes it hard for models to generalize across different clinics.
- The "Overconfidence" Trap: Standard neural networks often produce high-confidence scores even when they are wrong. In a clinical setting, a model that doesn't know when to say "I’m not sure" is a liability.
Methods
The authors propose CLEAR-Mamba, which builds upon the MedMamba backbone—a medical adaptation of the Mamba architecture. Mamba uses Structured State Space models (SSMs) to capture long-range dependencies with linear complexity, making it much more efficient than Transformers for sequential data.
The framework introduces two novel components:
- HaC (Hyper-adaptive Conditioning): This is a hypernetwork-based layer that dynamically generates model parameters based on the specific distribution of the input features. Think of it as a "smart filter" that tunes itself on-the-fly to account for different camera types or patient characteristics, improving cross-domain adaptability.
- RaP (Reliability-aware Prediction): Instead of using a standard Softmax output (which just gives a point estimate), RaP employs Evidential Deep Learning (EDL). It treats the model's output as parameters of a Dirichlet distribution. This allows the model to quantify "epistemic uncertainty"—essentially measuring how much evidence it has for a prediction. If the evidence is low, the model signals high uncertainty, allowing for a "risk-aware deferral" to a human expert.
Datasets
The study utilizes a mix of a massive new private dataset and three public benchmarks:
- In-house Angiography Dataset: A large-scale collection of 15,524 images covering 43 ocular disease categories (e.g., Diabetic Retinopathy, CSC, Uveitis) plus healthy controls. This is notable for its "long-tailed" nature, reflecting real-world clinical rarity.
- RetinaMNIST: A subset of MedMNIST for diabetic retinopathy grading.
- Harvard-GDP: A dataset for glaucoma progression forecasting.
- OCT-C8: A multi-class retinal OCT benchmark.
To process the raw clinical data, the authors developed a Multi-Agent Data Engine that automatically extracts images from PDF reports, anonymizes them, and aligns them with clinical text.
Key Insights and Ideas
- Mamba is a Natural Fit for Angiography: Because angiography is inherently temporal, Mamba’s ability to handle sequences efficiently allows it to outperform CNNs (which have limited receptive fields) and Transformers (which are computationally expensive for long sequences).
- Adaptation over Fine-tuning: The HaC module suggests that we don't necessarily need to fine-tune models for every new device. Instead, we can build architectures that adapt their own weights based on the input data's "style."
- Calibration is as Important as Accuracy: The "aha moment" here is that the RaP module doesn't just make the model more accurate; it makes the confidence scores meaningful. A correctly classified sample typically shows much higher "evidence" than an incorrectly classified one, providing a safety mechanism for clinical deployment.
Key Results
- Superior Performance: On the 43-class in-house dataset, the "Base" version of CLEAR-Mamba achieved an Overall Accuracy (OA) of 59.06%, significantly higher than the original MedMamba-B (55.38%) and ResNet50 (51.47%).
- State-of-the-Art on Public Benchmarks: On the OCT-C8 dataset, CLEAR-S reached 94.5% OA, surpassing established CNN and hybrid models.
- Outperforming Multimodal Models: On the Harvard-GDP benchmark, CLEAR-Mamba achieved an accuracy of 0.91, outperforming several models that had access to multiple imaging modalities, despite CLEAR using only a single modality.
- Reliability: Visualization via t-SNE showed that CLEAR-Mamba creates much tighter, more distinct clusters for different diseases compared to ResNet or MambaVision, indicating better feature discriminability.
Limitations and Future Work
- Single-Modality Focus: While the model performs exceptionally well on single-modality angiography, the authors acknowledge that clinical diagnosis often involves combining FFA/ICGA with OCT or fundus photos. Future versions aim to integrate multimodal fusion.
- Prospective Testing: The results are currently based on retrospective datasets. The next step is multi-center prospective validation to see how the model handles real-time clinical workflows.
- Pre-training: The authors suggest that "sequence-aware" pre-training (specifically designed for the temporal nature of medical videos) could further boost the model's understanding of disease progression.
Summary updated