🤖 AI

Agentic LLM Workflow for MR Spectroscopy Volume-of-Interest Placements in Brain Tumors

This paper proposes an agentic large language model workflow that generates diverse candidate volumes-of-interest using vision transformer models with varying objectives and selects the optimal placement based on quantitative metrics, thereby reducing inter-operator variability and adapting magnetic resonance spectroscopy to specific clinical priorities for brain tumors without requiring task-specific model retraining.

Original authors: Sangyoon Lee, Francesca Branzoli, Małgorzata Marjańska, Patrick Bolan

Published 2026-03-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sangyoon Lee, Francesca Branzoli, Małgorzata Marjańska, Patrick Bolan

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 a chef trying to taste a very complex, multi-layered cake to figure out exactly what's inside. You have a small spoon (the Volume-of-Interest, or VOI) that you need to dip into the cake.

If you dip the spoon into the frosting, you only taste sugar. If you dip it into the burnt bottom, you taste charcoal. To get a true flavor profile of the whole cake, you need to place that spoon in the perfect spot: covering the rich chocolate layers, avoiding the burnt bits, and not touching the plate (the skull).

This is exactly the challenge doctors face with Magnetic Resonance Spectroscopy (MRS) for brain tumors. They need to place a tiny "measurement box" inside a patient's brain to analyze the chemistry of the tumor. But brains are messy, and tumors are like uneven cakes with hard lumps, soft mush, and dead spots.

The Problem: The "Human Guessing Game"

Traditionally, a human expert has to manually draw this box on a 3D image. It's a high-stakes game of "Goldilocks":

Too small? You miss the important parts of the tumor.
Too big? You include dead tissue (necrosis) or fluid, which ruins the data.
Wrong angle? You might hit the skull, which messes up the magnetic signal.

Because every doctor has a slightly different style, and every tumor is unique, two experts might draw two completely different boxes for the same patient. This leads to inconsistent results.

The Old Solution: The "One-Size-Fits-All" Robot

Scientists tried to build AI robots to do this automatically. But most robots are like a strict recipe book: they are trained to do one specific thing perfectly. If you want the robot to focus on the "chocolate" (solid tumor) instead of the "frosting" (periphery), you usually have to throw away the old robot and train a brand new one from scratch. That's slow, expensive, and impractical.

The New Solution: The "Agentic Chef" (The Paper's Innovation)

This paper introduces a new system that acts like a smart, adaptable kitchen manager using an Agentic Large Language Model (LLM). Think of it as a team of three specialists working together:

1. The "Diverse Generators" (The Creative Sous-Chefs)

Instead of one robot trying to guess the perfect spot, the system uses AI models to generate many different "good enough" options.

Imagine asking five different chefs to suggest where to put the spoon. One might say, "Put it right in the center!" Another might say, "Let's aim for the edge to catch the sauce."
These models are trained to create a variety of valid boxes, not just one. Some are small and tight; some are big and loose.

2. The "Agentic Manager" (The LLM)

This is the brain of the operation. It's a Large Language Model (like a super-smart AI assistant) that acts as the project manager.

The User speaks: A doctor types a simple instruction: "I need to maximize the solid tumor coverage, even if the box gets a little bigger," or "I need to avoid the dead tissue at all costs."
The Manager thinks: The LLM reads this request, understands the goal, and looks at the list of "sous-chef" suggestions generated in step 1.
The Manager decides: It picks the best box from the list that matches the doctor's specific request. It doesn't need to be retrained; it just uses its "common sense" to choose the right tool for the job.

Why This is a Big Deal

Flexibility: In the past, if a doctor wanted a different type of measurement, they needed a different AI model. Now, they just change the words they type, and the system adapts instantly.
Better than Humans: On 110 real patient cases, this system was able to cover more of the actual tumor and avoid more dead tissue than the average human expert, while still keeping the box size reasonable.
No "Retraining" Headaches: You don't need to feed the AI thousands of new examples every time a new clinical question arises. You just tell it what you want, and it figures out which of its pre-made options fits best.

The Analogy in a Nutshell

Old Way: You have a single, rigid robot arm that can only place a box in one specific way. If you want it to do something else, you have to rebuild the robot.
New Way: You have a smart manager who asks a team of artists to draw 50 different boxes. You then tell the manager, "I want the one that covers the most chocolate." The manager instantly picks the perfect drawing from the pile.

This "Agentic Workflow" turns a rigid, technical problem into a flexible conversation, allowing doctors to get exactly the data they need for every unique patient, without the headache of constantly retraining their software.

1. Problem Statement

Magnetic Resonance Spectroscopy (MRS) is a critical tool for the non-invasive metabolic characterization of brain tumors, aiding in diagnosis, treatment planning, and monitoring. However, the clinical utility of MRS is heavily dependent on the accurate placement of the Volume-of-Interest (VOI).

Challenges:
- Complex Optimization: VOI placement requires balancing competing objectives: maximizing coverage of solid tumor tissue, avoiding necrosis and cerebrospinal fluid (which degrade spectral quality), minimizing partial-volume contamination, and ensuring adequate signal-to-noise ratio.
- Inter-operator Variability: Manual placement by experts varies significantly due to heterogeneous tumor subregions and differing clinical priorities.
- Multiple Valid Solutions: For a single tumor, multiple spatially distinct VOI placements can satisfy clinical constraints.
- Limitations of Current Automation: Traditional supervised learning or discrete optimization approaches typically encode a single, fixed objective. Adapting these models to different clinical questions or user preferences requires retraining separate models for each scenario, which is impractical.

2. Methodology

The authors propose an Agentic Large Language Model (LLM) workflow that decouples the problem into two stages: (1) generating diverse, acceptable VOI candidates, and (2) selecting the optimal candidate based on natural language user preferences.

A. Agentic Workflow Architecture

LLM Controller: The system uses Qwen3-14B as the central agent. It acts as a high-level controller that interprets user instructions (e.g., "maximize solid tumor coverage" or "minimize necrosis inclusion") and orchestrates a pipeline of tools.
Toolset: The LLM coordinates four specific tools:
1. Tumor Segmentation: Uses an nnUNet model trained on BraTS 2017 data to segment contrast-enhancing (CE), non-enhancing (NE), and necrotic regions.
2. VOI Placement: Generates candidate VOIs using deep learning models.
3. Metrics Evaluation: Quantifies the overlap between the VOI and specific tissue classes.
4. Task Completion: Finalizes the selection.

B. Candidate Generation Models

To address the "multiple valid solutions" problem, the authors developed two Vision Transformer (ViT)-based models:

MVP (MRS VOI Placement): A standard regression model that predicts a 9D vector (center, dimensions, rotation) directly from tumor segmentation maps. It is trained to produce a single deterministic placement based on a specific objective function.
PC-MVP (Position-Conditioned MVP): A variant that takes an additional coordinate input (intended VOI center). By sampling different positions across the solid tumor, PC-MVP generates a diverse set of acceptable VOIs (capped at 50 candidates) rather than a single deterministic output.

C. Data Augmentation via Discrete Search

To train these models without extensive manual annotation for every possible preference, the authors used a discrete objective search to generate synthetic training data:

Objective Function: A composite score based on Gaussian terms for tissue fractions (solid tumor, periphery, necrosis, normal brain) and geometric modifiers (penalties for missing solid tumor, excessive volume, or proximity to the skull).
Preference Sets: Two parameter sets were defined:
- Balanced: Reflects average expert behavior (equal weight on coverage and necrosis avoidance).
- Large-VOI / Low $f_{SolidOutside}$ : Emulates experts prioritizing maximum tumor coverage, even at the cost of larger volumes.
Training Data: 500 BraTS2021 cases were used to generate training datasets with these specific preferences.

D. LLM-Based Selection

The LLM receives the list of candidate VOIs generated by MVP/PC-MVP along with their quantitative metrics. Based on the user's natural language prompt (e.g., "Select the VOI with the lowest necrosis inclusion"), the LLM evaluates the metrics and selects the final optimal VOI.

3. Key Contributions

Objective-Driven Data Generation: A novel method to create supervised training datasets for different VOI placement preferences using a discrete search objective function, eliminating the need for massive manual annotation of diverse expert styles.
Diverse Candidate Generation: The introduction of PC-MVP, which produces multiple valid VOI candidates by conditioning on different spatial positions, enabling the system to explore the solution space rather than converging to a single point.
Agentic LLM Workflow: A tool-oriented pipeline where an LLM acts as a controller to interpret natural language preferences and select the best VOI from a pool of AI-generated candidates, allowing for dynamic adaptation to clinical goals without retraining.
Preference-Conditioned Control: Demonstration that the system can modulate VOI size and location to align with specific clinical objectives (e.g., maximizing coverage vs. minimizing necrosis) while maintaining expert-like placement patterns.

4. Results

The system was evaluated on 110 clinical brain tumor cases (gliomas) and compared against five human experts and previous automated methods.

Performance vs. Experts:
- The Agentic Workflow outperformed human experts and prior deep learning methods in specific metrics depending on the preference.
- Low Necrosis Preference: Achieved the lowest necrotic tissue inclusion ( $0.35 \pm 0.59$ mL) among all methods.
- High Coverage Preference: Achieved the highest solid tumor coverage ( $0.66 \pm 0.23$ ).
- Low Missed Tumor Preference: Achieved the lowest fraction of solid tumor outside the VOI ( $0.52 \pm 0.22$ ).
Volume Stability: Unlike discrete search methods that drastically altered VOI volume when optimizing for specific metrics, the agentic workflow maintained VOI volumes within a range comparable to human experts ( $\approx 12$ mL), ensuring clinical feasibility.
Flexibility: The workflow successfully reproduced distinct expert behaviors (e.g., one expert favoring small, core-targeted VOIs vs. another favoring large, comprehensive VOIs) simply by changing the natural language prompt, without retraining the underlying models.

5. Significance

This work represents a paradigm shift in medical image analysis automation:

Adaptability without Retraining: It solves the "one model, one task" limitation of supervised learning. By separating candidate generation from selection, the system can adapt to new clinical questions or institutional preferences via natural language prompts rather than retraining.
Handling Ambiguity: It acknowledges that medical imaging often has multiple "correct" answers. By generating a diverse set of candidates and using an LLM to select based on context, the system mimics the nuanced decision-making of human experts.
Clinical Integration: The workflow provides a practical strategy for integrating AI into MRS protocols, potentially reducing inter-operator variability and saving time while tailoring the scan to specific diagnostic needs (e.g., distinguishing tumor recurrence from radiation necrosis).

In conclusion, the proposed agentic LLM workflow offers a robust, flexible, and clinically relevant solution for automating MRS VOI placement, bridging the gap between rigid algorithmic optimization and the nuanced, preference-driven decisions of human radiologists.