🤖 AI

Progressive Learning with Anatomical Priors for Reliable Left Atrial Scar Segmentation from Late Gadolinium Enhancement MRI

This paper proposes a clinically inspired, three-stage progressive learning framework with an anatomy-aware loss function to improve the reliability and accuracy of left atrial scar segmentation from LGE MRI by embedding anatomical priors and diagnostic reasoning into the deep learning process.

Original authors: Jing Zhang, Bastien Bergere, Emilie Bollache, Jonas Leite, Mikaël Laredo, Alban Redheuil, Nadjia Kachenoura

Published 2026-03-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jing Zhang, Bastien Bergere, Emilie Bollache, Jonas Leite, Mikaël Laredo, Alban Redheuil, Nadjia Kachenoura

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

The Big Picture: Finding Tiny Scars in a Noisy Room

Imagine your heart's left atrium (the upper-left chamber) is a giant, thin-walled balloon. Sometimes, due to a condition called Atrial Fibrillation (AF), this balloon gets damaged, leaving behind tiny, invisible "scars" on the inside. These scars are bad news because they can cause the heart rhythm to go haywire again after surgery.

Doctors use a special MRI scan (called LGE) to try to see these scars. But here's the problem:

The scars are tiny: They are like trying to find a few grains of sand on a beach.
The picture is fuzzy: The MRI images are often low-contrast, making the scars look like ghosts rather than clear objects.
The map is messy: Even the expert doctors who draw the "maps" (annotations) sometimes make mistakes, marking spots that shouldn't be scars.

The goal of this paper is to teach a computer (AI) to find these tiny, ghostly scars reliably, even when the picture is blurry and the map is imperfect.

The Solution: A "Three-Step" Training Camp

Instead of throwing the AI into the deep end and telling it, "Find the scars!" (which usually fails), the authors created a Progressive Learning strategy. Think of this like training a new medical resident doctor. You wouldn't let them perform heart surgery on day one. You'd teach them step-by-step.

Step 1: Learn the Shape of the Balloon (Pre-training)

The Analogy: Imagine you are learning to draw a house. Before you try to draw the tiny cracks in the bricks (the scars), you first learn to draw the whole house perfectly.
What the AI does: The AI is shown thousands of MRI scans and asked only to find the big, easy-to-see Left Atrium (the balloon). It ignores the scars completely.
Why? This teaches the AI the "anatomy" of the heart. It learns where the walls are and what the shape looks like. It builds a solid foundation.

Step 2: The "Shadow" Game (Dual-Task Learning)

The Analogy: Now that the doctor knows the shape of the house, we ask them to find the cracks while keeping an eye on the house shape. But we add a rule: "Cracks can only exist on the walls, never floating in the air or inside the empty rooms."
What the AI does: The AI now tries to find the scars, but it has to do it while still remembering the shape of the atrium.
The Secret Sauce (Anatomical Priors): The authors added a special "rulebook" (a weighted loss function). If the AI guesses a scar is floating in the middle of the empty space (which is physically impossible), the rulebook says, "Nope, that's wrong." If it guesses a scar is on the wall, it gets a gold star. This stops the AI from getting confused by the messy "maps" (annotation errors) and forces it to be realistic.

Step 3: The Final Exam (Fine-Tuning)

The Analogy: The doctor is now ready for the final test. They stop worrying about drawing the whole house and focus 100% on finding the tiny cracks.
What the AI does: The AI takes everything it learned in Steps 1 and 2 and focuses purely on the scar segmentation. It fine-tunes its vision to catch those tiny, low-contrast spots.

Why This Matters: The "Physics" of the Data

The authors also realized that the data they had was scarce and sometimes "noisy" (imperfect). To fix this, they invented a Data Augmentation strategy.

The Analogy: Imagine you are practicing for a driving test, but you only have one empty parking lot. To get better, you don't just drive in that one spot. You simulate driving in the rain, in the fog, on a bumpy road, and with a slightly tilted steering wheel.
What they did: They artificially "distorted" the MRI images (making them blurrier, shifting the brightness, stretching them slightly) to teach the AI to recognize scars even when the image quality isn't perfect. This makes the AI much tougher and more reliable in the real world.

The Results: Did It Work?

The Outcome: The three-step method worked better than trying to do it all at once.
- The AI got very good at finding the big heart shape (94% accuracy).
- For the tiny scars, it improved its accuracy slightly but significantly (from 49% to 50% Dice score, and the edges of the scars were much sharper).
The Catch: Finding these scars is really hard. Even the best AI in the world struggles because the scars are so small and the images are so noisy. The authors admit their results aren't "perfect" yet, but their method is more reliable and anatomically correct than previous methods.

The Takeaway

This paper isn't just about a new computer algorithm; it's about changing how we teach computers.

Instead of forcing an AI to memorize a complex medical problem all at once, the authors taught it to think like a doctor:

Understand the big picture first.
Apply real-world rules (scars can't float in mid-air).
Then, focus on the tiny details.

By mimicking human clinical reasoning, the AI becomes a more trustworthy partner for doctors, helping them spot the invisible scars that could save a patient's life.

1. Problem Statement

The segmentation of Left Atrial (LA) scar tissue from Late Gadolinium Enhancement (LGE) MRI is critical for assessing Atrial Fibrillation (AF) severity and predicting recurrence after ablation. However, automated segmentation faces significant challenges:

Low Contrast & Small Size: Scar tissue is often thin, fragmented, and exhibits low contrast relative to healthy myocardium.
Class Imbalance: Scar voxels constitute a tiny fraction of the total volume (approx. 0.69% in the LASCARQS dataset).
Annotation Variability: Manual annotations often contain noise, including labels placed outside the anatomically plausible LA wall or inside the LA cavity.
Model Instability: Direct end-to-end training often leads to unstable convergence or failure to capture fine-grained pathological details due to the complexity of the task.

2. Methodology

The authors propose a 3-stage progressive learning framework based on the SwinUNETR architecture, inspired by the clinical diagnostic workflow where a physician first identifies the anatomy before focusing on pathology.

A. Data Augmentation & Loss Function

Augmentation: To address limited data and domain shifts, a physics-inspired, on-the-fly 3D augmentation pipeline was used. It includes bias field correction, 3D elastic deformation, rotation, zoom, blurring, contrast adjustment, intensity shifts, and Gaussian noise.
Compound Loss: A combination of Dice Loss and Cross-Entropy (DiceCE) is used to handle severe class imbalance.

B. The Three-Stage Progressive Framework

Stage I: Pre-training on LA Cavity Segmentation
- Goal: Establish robust anatomical priors.
- Process: The model is trained only to segment the LA cavity (ignoring scars). This allows the encoder to learn stable cardiac anatomy without the noise of scar detection.
- Outcome: A pre-trained encoder and LA decoder.
Stage II: Dual-Task Learning with Anatomical Priors
- Goal: Learn the spatial relationship between LA geometry and scar patterns.
- Architecture: A dual-decoder design. The encoder and LA decoder inherit weights from Stage I; a new scar decoder is randomly initialized.
- Anatomy-Aware Spatially Weighted Loss: A novel contribution where a "soft wall constraint" is applied. A spatial weight map ( $W$ $W$ ) is generated based on the distance from the LA boundaries.
  - Predictions outside the plausible LA wall region are down-weighted.
  - This mitigates annotation bias (e.g., false positives inside the cavity) and enforces anatomical plausibility without altering the fundamental loss function.
- Optimization: Joint optimization of LA and scar tasks, with a weighting bias toward scar ( $\beta_{scar}=0.7$ ).
Stage III: Fine-Tuning on Scar Segmentation
- Goal: Specialize the model for precise scar delineation.
- Process: The network is simplified to focus exclusively on the scar decoder. The early encoder layers are frozen to preserve low-level anatomical features learned in Stage I, preventing catastrophic forgetting.
- Outcome: Final fine-tuning on the scar task using the anatomical priors established in previous stages.

3. Key Contributions

Clinically Inspired Progressive Learning: A novel training strategy that mimics clinical reasoning (Anatomy $\to$ Context $\to$ Pathology) to stabilize learning for low-contrast, imbalanced tasks.
Anatomy-Aware Spatially Weighted Loss: A mechanism that incorporates clinical prior knowledge (scars exist only in the wall) to suppress anatomically implausible predictions and reduce sensitivity to annotation noise.
Physics-Informed Augmentation: A tailored 3D augmentation strategy specifically designed for LGE MRI characteristics to improve generalization.

4. Results

The method was evaluated on the LASCARQS 2022 public dataset (Subset 1: 60 volumes with scar annotations) using 5-fold cross-validation.

LA Segmentation: Achieved a Dice Score (Dsc) of 0.94 in Stage I. Performance slightly decreased to 0.90 in later stages due to task interference but remained robust.
Scar Segmentation:
- Stage I+II (Dual-task): Dsc = 0.49, Hausdorff Distance (HD) = 13.02 mm, Average Surface Distance (ASD) = 1.96 mm.
- Stage I+II+III (Full Framework): Dsc = 0.50, HD = 11.84 mm, ASD = 1.80 mm.
- The full framework outperformed the dual-task model, showing a 2% improvement in Dice and reductions in HD (9%) and ASD (8%).
Ablation Study:
- Direct supervised learning (Stage III only) yielded Dsc = 0.49.
- Adding Stage I pre-training (B2) improved Dsc to 0.50 and significantly reduced HD/ASD, confirming the value of anatomical pre-training.
- The full 3-stage approach provided the most consistent results, though the absolute Dice score remains modest due to the extreme difficulty of the task.
Comparison: The approach performed comparably to the strong baseline nnU-Net (Dsc 0.50) but with better surface distance metrics, highlighting the benefit of the progressive strategy over standard architectures.

5. Significance and Conclusion

Clinical Relevance: The study demonstrates that embedding clinical reasoning (anatomical constraints) into deep learning architectures improves the reliability of scar segmentation, even when quantitative metrics (like Dice) are limited by the inherent difficulty of the task.
Robustness to Noise: The proposed weighted loss function effectively handles the "label noise" common in public datasets where annotations may violate anatomical rules.
Future Outlook: While the Dice score improvement over state-of-the-art baselines is marginal, the reduction in surface distance errors and the improved anatomical consistency suggest the model is more clinically viable. The authors emphasize that for scar segmentation, regional interpretation may be more clinically relevant than strict voxel-wise overlap.

In summary, this paper presents a robust framework that bridges the gap between clinical workflow and AI modeling, proving that progressive learning and anatomical priors are essential for solving high-difficulty medical image segmentation tasks characterized by extreme class imbalance and annotation noise.