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

The paper introduces CLEAR-Mamba, an enhanced MedMamba framework featuring a hypernetwork-based adaptive conditioning layer and a reliability-aware prediction scheme, which achieves superior accuracy and trustworthiness in multi-sequence ophthalmic angiography classification by addressing challenges in generalization and confidence estimation.

The Gist

Imagine you are a detective trying to solve a mystery inside a patient's eye. The clues aren't just a single photo; they are a movie (a sequence of images) showing how blood flows and how lesions (damaged areas) change over time. This is what doctors call Ophthalmic Angiography.

For a long time, computer programs (AI) trying to read these eye movies have struggled. They often get confused by the subtle changes, they get "overconfident" when they are actually wrong, and they fail when they see a new type of camera or a rare disease they haven't seen before.

Enter CLEAR-Mamba. Think of it as a brand-new, super-smart detective agency designed specifically to solve these eye mysteries. Here is how it works, broken down into simple parts:

1. The Problem: The "Static Photo" Trap

Most old AI models treat a movie like a stack of still photos. They look at one frame and guess the disease.

• The Flaw: It's like trying to understand a soccer game by looking at a single frozen photo of the players. You miss the movement, the strategy, and the flow.
• The Result: The AI misses the "story" of the disease (how it grows or leaks fluid over time).

2. The Solution: The "Mamba" Backbone

CLEAR-Mamba uses a special engine called MedMamba.

• The Analogy: Imagine a traditional AI (like a CNN) is a person reading a book one word at a time, forgetting the beginning by the time they reach the end.
• The Mamba: This is like a person who can hold the entire story in their head while reading. It can look at the first frame of the eye movie and the last frame simultaneously, understanding the flow of time and blood. It's fast, efficient, and remembers the whole sequence.

3. The "Chameleon" Layer (HaC)

One of the biggest headaches in medical AI is that every hospital uses different cameras, and every patient's eyes look slightly different. A model trained on one camera often fails on another.

• The Analogy: Imagine a detective who only knows how to speak English. If they go to a country where everyone speaks French, they are useless.
• The HaC (Hyper-adaptive Conditioning): This is like giving the detective a magic translator that instantly learns the local dialect. Before the detective looks at the clues, this layer "chameleon-izes" itself to match the specific camera and patient. It adjusts its own internal rules on the fly so it can understand any eye movie, no matter where it came from.

4. The "Honesty" Module (RaP)

Old AI models are terrible at admitting when they are confused. If you show them a blurry, weird image, they will still give you a 99% confident answer, which is dangerous for a doctor.

• The Analogy: Imagine a weather forecaster who says "100% chance of rain" even when the sky is clear. You can't trust them.
• The RaP (Reliability-aware Prediction): This module teaches the AI to be honest. It doesn't just guess the disease; it calculates how much "evidence" it has.
  • High Evidence: "I'm 95% sure this is Glaucoma." (The doctor trusts this).
  • Low Evidence: "I'm only 40% sure. This looks weird. Please, human doctor, take a second look."
  • This prevents the AI from making confident mistakes. It knows when to say, "I don't know."

5. The Massive Training Ground (The Dataset)

To teach this detective, the researchers didn't just use a few photos. They built a massive library of 43 different eye diseases using thousands of real-world eye movies (FFA and ICGA scans).

• The Challenge: Real medical data is messy. Some diseases are common (like Diabetes in the eye), and some are super rare. The data is also full of "noise" (blurry images, patient names hidden, etc.).
• The Fix: They built a robot team (an automated pipeline) to clean, organize, and label this messy data so the AI could learn properly.

The Result: Why This Matters

When they tested CLEAR-Mamba:

1. It was more accurate: It caught diseases better than older models (ResNet, ViT, and even the original MedMamba).
2. It was more trustworthy: It knew when to be unsure, which is crucial for patient safety.
3. It was adaptable: It worked well even on data it had never seen before (like different hospitals or different types of eye scans).

In a Nutshell

CLEAR-Mamba is like upgrading a detective from a rookie who only looks at still photos and guesses wildly, to a seasoned veteran who:

• Watches the whole movie to understand the story.
• Instantly adapts to new languages and accents.
• Is brave enough to say, "I'm not sure, let's ask a human," rather than making a dangerous guess.

This makes it a huge step forward for using AI to help doctors diagnose eye diseases earlier and more safely.

Technical Summary

Here is a detailed technical summary of the paper "CLEAR-Mamba: Towards Accurate, Adaptive and Trustworthy Multi-Sequence Ophthalmic Angiography Classification."

1. Problem Statement

The paper addresses critical limitations in current Computer-Aided Diagnosis (CAD) systems for ophthalmic angiography, specifically Fluorescein Fundus Angiography (FFA) and Indocyanine Green Angiography (ICGA). Key challenges include:

• Underutilization of Temporal Dynamics: Angiography is inherently a time-series data capturing hemodynamic changes (e.g., early filling to late leakage). Most existing methods treat these sequences as static images or focus heavily on multimodal fusion (combining FFA with OCT/CFR), neglecting the rich temporal information within a single modality.
• Architectural Limitations: Traditional CNNs struggle with long-range temporal dependencies due to limited receptive fields, while Vision Transformers (ViTs) are computationally expensive and memory-intensive, hindering real-time clinical deployment.
• Lack of Reliability and Adaptability:
  • Reliability: Standard softmax outputs are often overconfident on noisy or out-of-distribution data, failing to distinguish between epistemic (model) and aleatoric (data) uncertainty.
  • Adaptability: Models often degrade when applied to multi-disease settings or across different devices due to domain shifts. Existing adaptation methods rarely account for the specific temporal progression patterns of ophthalmic lesions.
• Data Scarcity: There is a lack of large-scale, high-quality, single-modality datasets covering diverse retinal diseases with complete temporal sequences.

2. Methodology: CLEAR-Mamba

The authors propose CLEAR-Mamba, a framework designed to be accurate, adaptive, and trustworthy. It integrates three core components:

A. Backbone: MedMamba

The framework utilizes MedMamba, which employs Visual State Space Models (VSSMs) and the 2D Selective Scan (SS2D) mechanism.

• Function: Efficiently captures both fine-grained local vascular textures and long-range global structural dependencies with linear computational complexity, making it suitable for high-resolution angiography sequences.

B. Adaptive Module: HaC (Hyper-adaptive Conditioning)

To address domain shifts and case-specific variability, the authors introduce a HyperNetwork-based Adaptive Conditioning layer.

• Mechanism: A lightweight hypernetwork takes input features ($z$) and dynamically generates parameters (scaling $\gamma$ and shifting $\beta$ factors) for the main backbone.
• Implementation: It uses a FiLM (Feature-wise Linear Modulation) approach combined with a conditioned gate to stabilize updates. This allows the model to adapt its feature extraction strategy on a per-sample basis without heavy fine-tuning, improving cross-domain generalization.

C. Prediction Head: RaP (Reliability-aware Prediction)

To ensure trustworthy decision-making, the framework replaces standard classification heads with an Evidential Learning approach.

• Mechanism: Instead of outputting deterministic logits, the head predicts the parameters ($\alpha$) of a Dirichlet distribution.
• Output: This yields both class probabilities and a measure of uncertainty (predictive entropy and total evidence).
• Training: The model is trained using a marginal likelihood loss combined with a Kullback-Leibler (KL) divergence regularizer toward a non-informative prior. This encourages the model to express high uncertainty when evidence is insufficient, enabling "risk-aware deferral" (i.e., flagging ambiguous cases for human review).

3. Key Contributions

1. Methodological Innovation (CLEAR-Mamba):

   • A unified framework combining efficient temporal modeling (Mamba), lightweight instance-wise adaptation (HaC), and calibrated uncertainty estimation (RaP).
   • Demonstrates that single-modality temporal modeling can outperform complex multimodal fusion approaches when temporal dynamics are properly leveraged.

2. Dataset Curation:

   • Developed a large-scale, single-modality, multi-sequence dataset containing 15,524 valid images covering 43 distinct ocular disease categories (plus healthy controls) from FFA and ICGA modalities.
   • Created an automated multi-agent pipeline to extract, anonymize, align, and clean raw PDF clinical reports, solving issues of privacy, label noise, and modality parsing.

3. Comprehensive Evaluation:

   • Validated the model on the in-house dataset and three public benchmarks (RetinaMNIST, OCT-C8, Harvard-GDP), showing consistent superiority over CNN, ViT, and Mamba baselines.

4. Experimental Results

• In-House Dataset (43 Diseases):

  • CLEAR-Mamba significantly outperformed strong baselines (ResNet, DINOv3, MedViT, MedMamba).
  • Performance: The CLEAR-B variant achieved 59.06% Overall Accuracy (OA) and 22.71% F1-score, improving upon the best baseline (MedMamba-X) by ~4% in OA and ~6-8% in F1.
  • Reliability: Ablation studies showed that HaC and RaP together improved confidence calibration. The full model achieved better separation between correct and incorrect predictions (median confidence gap) compared to models with only one component.

• Public Benchmarks:

  • OCT-C8: CLEAR-S achieved 94.5% OA and 0.9961 AUC, surpassing all CNN and hybrid medical vision baselines.
  • RetinaMNIST: CLEAR-B achieved 56.8% OA, outperforming MedMamba and AutoML baselines.
  • Harvard-GDP: CLEAR achieved 0.91 Accuracy on glaucoma progression forecasting, outperforming reported multimodal fusion models using only single-modality (OCT) input.

• Analysis:

  • t-SNE Visualization: CLEAR embeddings showed tighter intra-class clustering and clearer inter-class separation compared to other architectures.
  • Uncertainty Quantification: Case studies demonstrated that the model correctly identified ambiguous cases (e.g., uveitis vs. hemorrhage) with high uncertainty, whereas baselines often made overconfident errors.

5. Significance and Impact

• Clinical Trustworthiness: By integrating evidential learning, CLEAR-Mamba moves beyond simple classification to provide confidence-aware predictions. This is crucial for high-stakes medical settings where knowing when the model is unsure is as important as the diagnosis itself.
• Efficiency and Generalization: The use of Mamba and HyperNetworks offers a computationally efficient solution that adapts to diverse clinical scenarios without requiring massive multimodal data, making it more deployable in resource-constrained environments.
• Data Resource: The release of a curated, multi-disease angiography dataset fills a significant gap in ophthalmic AI research, enabling future studies on temporal dynamics and rare diseases.
• Paradigm Shift: The study challenges the notion that multimodal fusion is always necessary, demonstrating that deep temporal modeling of single-modality angiography can yield state-of-the-art results for complex, multi-disease classification tasks.