LinGuinE: Longitudinal Guidance Estimation for Volumetric Tumour Segmentation

LinGuinE is a novel, training-free PyTorch framework that achieves state-of-the-art longitudinal volumetric tumour segmentation and lesion tracking across multiple datasets by combining image registration with guided segmentation from a single radiologist prompt, enabling flexible, direction-agnostic analysis without requiring longitudinal data training.

The Gist

Imagine you are a doctor trying to track a tumor in a patient's body over several months. You have a series of CT scans, like a photo album of the patient's insides, taken at different times. Your goal is to draw a precise outline around the tumor in every single photo to see if it's shrinking (good) or growing (bad).

Doing this manually for dozens of photos is exhausting and prone to human error. Doing it automatically with AI is tricky because the tumor moves, changes shape, and the patient's body shifts slightly between scans. Most current AI tools are like "single-photo artists"—they are great at drawing the tumor in one picture, but they get lost when you show them the next picture in the album. They don't know, "Hey, that blob in photo #2 is the same tumor from photo #1."

Enter LinGuinE (Longitudinal Guidance Estimation). Think of it as a smart, time-traveling GPS for tumors.

The Core Problem: The "Lost in Translation" Issue

Imagine you have a map of a city (Scan A) where you've marked a specific coffee shop (the tumor). You want to find that same coffee shop on a new map of the city taken a month later (Scan B). The city has changed: new buildings are up, roads are shifted, and the coffee shop might have moved or changed size.

• Old AI methods try to guess where the coffee shop is on the new map from scratch. They often guess wrong or lose track of it entirely.
• LinGuinE takes your original mark, uses a "magic ruler" (image registration) to stretch and align the old map onto the new one, and drops a pin exactly where the coffee shop should be. Then, it asks a specialist to refine that pin into a perfect outline.

How LinGuinE Works (The "GPS" Analogy)

1. The Starting Point (The Prompt):
   A radiologist (the expert) looks at just one scan (the "Source") and clicks on the tumor. This is like telling the GPS: "We are starting here."

   • Why this matters: The doctor doesn't have to click on every single scan. Just one click sets the whole journey in motion.

2. The Time Travel (Image Registration):
   The system uses a "shape-shifting" tool to warp the Source scan so it perfectly matches the shape and position of the next scan (the "Destination"). It's like taking a transparent sheet with your dot on it and stretching it until it fits perfectly over the new photo.

   • The Magic: It carries your dot from the old photo to the new one, even if the patient took a deep breath or the tumor moved slightly.

3. The Refinement (Guided Segmentation):
   Once the dot lands on the new scan, it might be slightly off-center because tumors change shape. LinGuinE then uses a smart AI assistant to look at that dot and say, "Okay, I see the tumor is actually here," and draws the perfect 3D outline.

   • The "Boost": Sometimes, the system does a "double-check." It draws a rough circle around the dot, finds the center of that circle, and uses that new center to draw the final, perfect outline. This is like adjusting your aim after a first shot.

Why LinGuinE is a Game-Changer

• No "Time Travel" Training Needed:
  Most AI models need to be trained on thousands of "before and after" photo pairs to learn how to track things. LinGuinE is different. It's like a universal adapter. You can plug in any existing "single-photo" AI tool and any "map-aligning" tool, and LinGuinE makes them work together for tracking. It doesn't need special training data to do this.

• Time is Flexible (Direction Agnostic):
  Usually, you have to track a tumor from the past to the future (Scan 1 $\to$ Scan 2 $\to$ Scan 3). LinGuinE is like a rewind button. A doctor can pick any scan in the middle of the series as the starting point and track forward or backward. This is huge for looking at old patient records where the "first" scan might be missing or unclear.

• It Doesn't Get Tired:
  The paper tested this on hundreds of patients over many weeks. While other methods started to lose accuracy the further apart the scans were (like a GPS signal getting weak over a long drive), LinGuinE stayed accurate. It barely lost any precision even after months of time separation.

The Bottom Line

LinGuinE is a plug-and-play toolkit that turns a single doctor's click into a complete, 3D movie of a tumor's life story. It combines the best existing tools for "aligning maps" and "drawing shapes" to solve a problem that was previously very hard: keeping track of a moving target over time without needing to retrain the AI from scratch.

It's not just about drawing lines; it's about giving doctors a reliable, automated way to see if a cancer treatment is actually working, turning a mountain of manual work into a simple, guided process.

Technical Summary

1. Problem Statement

Longitudinal volumetric tumour segmentation is essential for radiotherapy planning and assessing oncological treatment response. However, current deep learning approaches face three critical limitations:

• Single-Timepoint Focus: Most methods generate semantic masks for a single scan at a single timepoint, failing to track lesions across a series of scans.
• Lack of Lesion Correspondence: Fully automatic methods do not maintain correspondence between the same tumour across different timepoints, which is crucial for measuring progression or regression.
• Limited Radiologist Control: Existing fully automatic pipelines do not allow radiologists to select specific lesions of interest, reducing clinical adoption. Conversely, existing semi-automatic or tracking methods either lack segmentation capabilities, require extensive longitudinal training data (which is scarce), or are restricted to specific modalities/cancer types.

2. Methodology: The LinGuinE Framework

LinGuinE (Longitudinal Guidance Estimation) is a PyTorch framework designed to perform lesion-level tracking and volumetric segmentation across an entire longitudinal study using a single radiologist prompt provided at any single timepoint.

Core Workflow

The framework operates by combining image registration with guided segmentation:

1. Input: A source scan ($I_i$) and a prompt (e.g., a point $P^k_i$ or a mask) identifying a specific tumour $k$.
2. Point Propagation: The prompt coordinates are propagated from the source scan ($I_i$) to all destination scans ($I_j$) in the study using an image registration algorithm. The propagated point is denoted as $P^k_{i \to j}$.
   • Assumption: Tumours maintain anatomically similar locations across timepoints.
   • Flexibility: The framework is modular, allowing the use of any registration algorithm (e.g., GradICON, LungGradICON).
3. Guided Segmentation: The propagated point serves as a prompt for a semi-automatic segmentation model (e.g., nnInteractive, Vista3D, LesionLocator) to generate a mask $M_j(k)$ on the destination scan.
4. Post-Processing: To avoid false positives in nearby tumours, the system retains the connected component closest to the propagated point.

Advanced Features

• Prompt Boosting: An optional step where the center of an initial segmentation (generated by the propagated point) is used as a refined prompt for a second segmentation pass. This helps correct cases where the propagated point lands slightly outside the tumour due to shape/size changes.
• Temporal Direction Agnosticism: Unlike autoregressive methods that propagate sequentially (Scan 1 $\to$ 2 $\to$ 3), LinGuinE propagates directly from the source scan to any other scan in the series. This allows radiologists to choose the most informative scan as the source, regardless of whether it is the earliest, middle, or latest scan.
• No Longitudinal Training Required: The framework repurposes existing point-prompted segmentation models and registration algorithms without requiring them to be trained on longitudinal data.

3. Key Contributions

1. State-of-the-Art Performance: Achieves superior segmentation and tracking results across four diverse datasets with 456 longitudinal studies, outperforming existing methods like LesionLocator and Hering et al.
2. Temporal Agnosticism: The first framework allowing radiologists to provide prompts at any timepoint, facilitating retrospective analysis and reducing temporal degradation.
3. Modularity and Customization: Allows the integration of arbitrary registration and segmentation algorithms. The authors demonstrated that pathology-specific fine-tuning (e.g., for lung cancer) further enhances performance.
4. Public Benchmarking: Released a PyTorch package and the first public benchmark for longitudinal segmentation across multiple cancer types and timepoints, addressing a gap in open-source tools.

4. Experimental Results

The authors evaluated LinGuinE on four datasets: 4DCBCT (NSCLC), Phase-3 (Private NSCLC), autoPET Longitudinal CT (Melanoma), and UniToChest (Lung Cancer).

• Segmentation Accuracy (Dice Score): LinGuinE configurations consistently outperformed the pretrained LesionLocator model and the Hering et al. method (nnUNet-based) across all datasets.
  • Best Configuration: For lung datasets, combining LungGradICON (LGI) registration with fine-tuned LesionLocator Segmentation (LLSeg) yielded the highest Dice scores (e.g., 0.786 on UniToChest).
  • Generalization: Even without fine-tuning, using general-purpose algorithms (UGI + LLSeg) outperformed specialized methods trained on longitudinal data.
• Tracking Accuracy (Euclidean Distance): LinGuinE maintained lower center-point error (ED) compared to competitors, particularly in lung datasets.
• Temporal Degradation:
  • LinGuinE showed minimal performance degradation over time (mean decline of 0.126% per week).
  • In contrast, LesionLocator declined by 2.591% per week, and Hering et al. by 0.457% per week.
  • Statistical tests confirmed no significant difference in performance whether the prompt originated from the earliest, middle, or latest scan.
• Ablation Studies:
  • Radiologist Input: Using a single click yielded performance (Dice 0.806) nearly identical to inter-radiologist manual segmentation (0.810), though full masks as input provided a statistically significant slight improvement.
  • Boosting: Improved Euclidean Distance (tracking) in 3/4 datasets, particularly helpful when registration struggles.
  • Auto-regression: Contrary to LesionLocator's approach, LinGuinE's non-autoregressive (direct propagation) approach performed better or equally well, avoiding the compounding of errors over time.

5. Significance and Conclusion

LinGuinE addresses a critical gap in medical imaging by providing a robust, out-of-the-box solution for longitudinal tumour tracking without the need for scarce longitudinal training data.

• Clinical Impact: It enables precise measurement of treatment response and disease progression by maintaining lesion correspondence over time, a capability previously limited to manual or semi-automated workflows.
• Research Impact: By releasing the code and benchmark data, the authors facilitate future research into longitudinal analysis, moving the field away from single-timepoint evaluations.
• Future Work: The authors note that handling disappearing or coalescing lesions and extending the framework to other cancer types and modalities are necessary next steps.

In summary, LinGuinE represents a paradigm shift from "single-scan segmentation" to "longitudinal study analysis," leveraging the synergy of image registration and promptable segmentation to deliver clinically viable, high-accuracy tumour tracking.