💻 computer science

CoRE: Concept-Reasoning Expansion for Continual Brain Lesion Segmentation

The paper proposes CoRE, a Concept-Reasoning Expansion framework that integrates visual features with a hierarchical concept library to simulate clinical reasoning, thereby achieving state-of-the-art continual brain lesion segmentation with improved interpretability and efficient knowledge reuse across diverse MRI tasks.

Original authors: Qianqian Chen, Anglin Liu, Jingyang Zhang, Yudong Zhang

Published 2026-04-29

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Qianqian Chen, Anglin Liu, Jingyang Zhang, Yudong Zhang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are training a new radiologist to spot brain tumors on MRI scans. In the real world, this doctor can't just memorize one type of tumor and stop; they need to learn about strokes, multiple sclerosis, and different types of cancer as they appear, one after another. The biggest problem is that if they study too hard on the new cases, they might forget everything they learned about the old ones. This is called "catastrophic forgetting."

Current AI models try to solve this by either:

Keeping a fixed team of experts: They have a set number of "specialists" (like a team of 5 doctors) who share the work. But if a new, weird disease shows up, the team is too small to handle it, and they get overwhelmed.
Hiring a new specialist for every single job: This works well, but soon you have a hospital with thousands of doctors, which is expensive and messy. Plus, they don't share notes, so they keep reinventing the wheel.

Enter CoRE (Concept-Reasoning Expansion).

The authors of this paper propose a smarter way to train this AI doctor. Instead of just looking at the blurry pixels of an MRI scan (image perception), CoRE teaches the AI to think like a human radiologist using structured concepts.

Here is how CoRE works, broken down into three simple steps:

1. The "Medical Textbook" (Brain Lesion Concept Library)

Before the AI starts learning, the researchers give it a "textbook" built by asking a super-smart AI (a Large Language Model) to describe brain diseases.

The Analogy: Imagine instead of just showing the AI a picture of a tumor, you give it a list of rules like: "Tumors often look like bright spots on T2 scans," or "Strokes often have a wedge shape."
This creates a library of concepts (like "edema," "ring enhancement," or "mass effect"). These are the stable, logical building blocks of medical knowledge, which don't change even if the MRI machine changes.

2. The "Smart Receptionist" (Concept-Guided Calibration)

When a new MRI scan comes in, the AI doesn't just guess which "expert" (a small part of the neural network) should look at it.

The Analogy: Think of the AI as a busy hospital. The Concept-Guided Calibration is a smart receptionist.
The receptionist looks at the patient's scan (the visual image) and checks the medical textbook (the concepts).
If the scan shows a "ring-shaped enhancement," the receptionist says, "Ah, this matches the 'Ring Enhancement' concept! Let's send this to Expert #7, who specializes in that."
This ensures the AI routes the task to the right expert based on meaning, not just how the picture looks. This prevents the AI from getting confused when the image quality changes.

3. The "Hiring Manager" (Concept-Driven Expansion)

What happens if a new type of disease appears that no one has seen before?

The Analogy: The Concept-Driven Expansion module acts as a hiring manager. It asks two questions:
1. Does the visual image look weird? (Image signal)
2. Does the medical concept match something we've never seen? (Concept signal)
The Rule: The AI only hires a new expert if both conditions are true. If the image looks weird but the concept is something we already know (like a common stroke), the AI just uses an existing expert.
This stops the AI from hiring thousands of unnecessary experts just because the lighting on a scan changed slightly. It only grows when it truly needs new knowledge.

Why is this a big deal?

The researchers tested CoRE on 12 different brain lesion tasks (like different types of tumors and strokes) in a row.

The Result: CoRE learned all 12 tasks better than any other method. It didn't forget the first tasks while learning the last ones.
The "Few-Shot" Superpower: When they gave CoRE very little data to learn a new task (like showing it only 10% of the usual pictures), it still performed amazingly well. This is because it leaned on its "medical textbook" (concepts) to fill in the gaps, whereas other AIs just panicked and failed.

In Summary

CoRE is like teaching an AI to be a doctor who reasons rather than just memorizes. By combining visual images with a structured library of medical concepts, it knows exactly which "expert" to use and when to grow a new one. This makes it smarter, more efficient, and much better at remembering what it learned in the past.

1. Problem Statement

Accurate brain lesion segmentation in MRI is critical for clinical diagnosis but faces significant challenges in Continual Learning (CL) scenarios.

Data Constraints: Strict privacy regulations and high annotation costs prevent the retention of historical data, necessitating models that learn sequentially without access to past datasets.
Task Heterogeneity: Brain imaging involves diverse modalities (e.g., T1w, FLAIR, DWI) and pathological variations (e.g., tumors, strokes, MS lesions).
Limitations of Existing CL Paradigms:
- Fixed Expert Pools: Pre-defined capacity limits the model's ability to handle long data streams.
- Task-Specific Expansion: Adding dedicated experts for every task leads to linear parameter growth and hinders knowledge sharing.
- Image-Perception Strategies: Current dynamic expansion methods rely solely on visual statistical shifts. These fail in medical imaging because tasks may look visually similar but differ in underlying pathophysiology, or look different but share the same clinical concept. This leads to redundant parameter expansion and catastrophic forgetting.

2. Methodology: The CoRE Framework

The authors propose Concept-Reasoning Expansion (CoRE), a framework that shifts from pure visual perception to structured clinical reasoning. It simulates a radiologist's cognitive process by decoupling visual features into invariant hierarchical concepts.

The framework consists of three core components built upon a Swin UNETR backbone with lightweight Adapter modules:

A. Brain Lesion Concept Library (BLC-Lib) Generation

Goal: Construct a structured knowledge base bridging clinical priors and visual representations.
Process: A two-stage pipeline leveraging Large Language Models (LLMs):
1. Theoretical Elicitation: Querying an LLM with templates to extract theoretical concepts regarding signal intensity, anatomical location, and morphology.
2. Visual Alignment: Using an advanced LLM (GPT-5) to analyze representative cases with ground-truth masks, refining theoretical concepts to match specific data distributions.
Output: A concept space embedding matrix ( $\hat{C}$ ) generated using the BiomedCLIP text encoder, representing hierarchical lesion attributes (e.g., "Ring Enhancement," "Perilesional Edema").

B. Concept-Guided Calibration (CGC)

Goal: Guide the routing of input data to specific adapter experts based on both visual and semantic cues.
Mechanism: A dual-routing mechanism:
1. Concept-Side Routing: Projects image tokens into the concept space to compute similarity with the BLC-Lib. This generates weights based on interpretable lesion attributes.
2. Image-Side Routing: Captures continuous distributional characteristics (visual statistics) that discrete concepts might miss.
3. Fusion: The final routing weight is a weighted sum ( $\lambda$ ) of both signals, ensuring expert selection is grounded in clinical logic while remaining sensitive to visual nuances.

C. Concept-Driven Expansion (CDE)

Goal: Dynamically determine when to add new adapters (architectural growth) to prevent redundancy.
Mechanism: A dual novelty detector that triggers expansion only if both conditions are met:
1. Concept-Side Signal: The routing confidence is low (flat distribution), indicating the input contains novel conceptual patterns not covered by existing experts.
2. Image-Side Signal: The reconstruction error of a lightweight autoencoder exceeds a threshold, indicating a significant visual distribution shift.
Benefit: This joint decision prevents the "linear parameter growth" seen in image-only methods, ensuring the model grows only when truly necessary.

3. Key Contributions

First CL Paradigm with Clinical Reasoning: CoRE is the first continual learning framework for brain lesion segmentation that integrates structured clinical concepts (via LLMs) to guide model evolution, moving beyond pure image-perception strategies.
Novel Architecture (CoRE): Introduces a synergistic system of BLC-Lib, CGC, and CDE that aligns visual tokens with a hierarchical concept space, enabling interpretable expert routing and demand-based model growth.
Efficiency and Stability: By grounding expansion decisions in clinical priors, CoRE significantly reduces redundant parameter expansion while maximizing knowledge reuse across diverse modalities and pathologies.
Superior Few-Shot Transferability: The framework demonstrates the ability to adapt to new tasks with limited data by leveraging accumulated conceptual knowledge.

4. Experimental Results

The authors evaluated CoRE on a sequential stream of 12 tasks across five datasets (BraTS, ATLAS, MSSEG, ISLES, WMH) and tested few-shot capabilities on BraTS2023-SSA.

Performance: CoRE achieved State-of-the-Art (SOTA) performance, outperforming existing methods (including L2P, Moe-adapters, CODA-P, and SEMA).
- Average Dice Similarity Coefficient (DSC): 79.90% (vs. 77.44% for the second-best).
- Backward Transfer (BWT): 98.96%, indicating near-perfect retention of knowledge on previous tasks.
Few-Shot Learning: In scenarios with only 10-30% of training data, CoRE significantly outperformed baselines, proving its ability to generalize from limited samples using conceptual priors.
Parameter Efficiency: Unlike image-perception methods that showed nearly linear adapter growth, CoRE stabilized at a compact architecture (e.g., 8–10 adapters for deep layers) by filtering out redundant expansions.
Interpretability: Visualization of routing weights showed that adapters converged on specific, clinically coherent concepts (e.g., specific tumor attributes), validating the "reasoning" aspect of the model.

5. Significance and Impact

Clinical Trustworthiness: By simulating radiologist reasoning, CoRE provides interpretable decisions, a crucial requirement for clinical AI adoption.
Scalability: The framework offers a sustainable solution for universal medical models that can adapt to evolving clinical tasks (new scanners, new diseases) without violating data privacy or requiring massive storage.
Generalizability: The success of using open-weight LLMs for concept generation suggests the framework can be deployed in secure, local clinical environments without reliance on proprietary APIs.
Paradigm Shift: The paper advocates for a shift in medical continual learning from "visual statistics" to "semantic/conceptual reasoning," setting a new direction for handling non-stationary medical data streams.