CorSeg-CineSAX: An Open-Source Deep Learning Framework for Fully Automatic Segmentation of Short-Axis Cine Cardiac MRI Across Multiple Cardiac Diseases

CorSeg-CineSAX is an open-source deep learning framework that utilizes a large-scale, multi-center dataset and an anatomically constrained post-processing pipeline to achieve robust, fully automatic segmentation of cardiac MRI across diverse and previously unseen disease categories with high clinical agreement.

The Gist

Imagine your heart is a complex, hard-working factory. To keep it running smoothly, doctors need to measure the size of its rooms (the chambers) and the thickness of its walls (the muscle). The gold standard for taking these measurements is a special movie camera called an MRI.

However, there's a problem: The movie is too long and too detailed for a human to watch and measure by hand.

A single heart MRI movie has hundreds of frames. A doctor has to sit down and manually trace the outline of the heart's walls on every single frame, like tracing a picture in a coloring book. This takes 15 to 45 minutes per patient, is boring, and different doctors might trace slightly different lines.

Enter CorSeg-CineSAX, a new "super-intelligent robot assistant" designed by researchers to do this tracing job instantly, perfectly, and for free.

Here is how it works, broken down into simple concepts:

1. The "Super-Student" (The Training Data)

Most AI models are like students who only study for a specific test using one textbook. If the test question changes slightly, they get confused.

• The Old Way: Previous AI tools were trained on tiny datasets (maybe 150 patients) from just one hospital. They knew how to handle "normal" hearts but got lost when they saw a heart with a rare disease or a weird shape.
• The CorSeg Way: The researchers built the largest library of heart movies ever created. They gathered data from 1,555 patients across 12 different hospitals in China, covering five different heart diseases.
• The Analogy: Instead of studying one textbook, this AI read 319,000 pages of heart movies. It saw every type of heart shape, every size, and every stage of the heartbeat. It learned that a heart isn't just a circle; it's a squishy, moving shape that changes every second.

2. The "One-Size-Fits-All" Strategy (The Input)

Many AI tools are picky. They might say, "I can only look at the heart if you crop the picture perfectly first," or "I need the whole 3D block of images at once." If the hospital's computer sends the data in a slightly different format, the AI crashes.

• The CorSeg Way: This tool is incredibly flexible. It looks at one slice of the heart at a time (like looking at a single slice of bread from a loaf), regardless of how the picture is framed or where the heart is sitting in the chest.
• The Analogy: Imagine a chef who can chop vegetables whether they are on a fancy cutting board, a paper plate, or a wooden table. CorSeg doesn't care about the "messy" real-world data; it just grabs the slice and gets to work.

3. The "Safety Inspector" (Anatomical Post-Processing)

Even the smartest AI can make silly mistakes. Sometimes, the AI might draw a tiny, floating piece of heart muscle that doesn't connect to anything (a "ghost" fragment), or it might draw the inner chamber of the heart poking through the outer wall (which is physically impossible).

• The CorSeg Way: After the AI draws the picture, a rule-based "Safety Inspector" runs a quick check. It has three simple rules:
  1. No Ghosts: If a piece of heart muscle is floating alone, delete it.
  2. No Poking: The inner chamber must be inside the wall, never touching the outside air.
  3. No Holes: The wall between the left and right sides must be solid, with no gaps.
• The Analogy: Think of the AI as a talented but impulsive artist. The Safety Inspector is the editor who says, "Great drawing, but you accidentally drew a hole in the wall. Let's fix that before we show it to the doctor." This step removed 100% of the impossible errors.

4. The Results: Why It Matters

The researchers tested this robot on hearts it had never seen before, including hearts from different countries (France, Germany, Spain) and different scanner machines.

• The Score: It got a 91% accuracy score (Dice Coefficient), which is basically as good as a human expert.
• The Magic: Even when it saw a disease it had never been trained on (like a specific type of heart failure), it still got it right. This is called "Zero-Shot Generalization"—it's like a student who learns math so well they can solve a problem they've never seen before.
• The Output: It calculated the heart's pumping power and volume almost instantly, matching human measurements perfectly.

The Big Picture

CorSeg-CineSAX is not just a new algorithm; it's a tool for the people.

• It's Open Source: The code is free for anyone to download and use.
• It's Robust: It works even if the data is messy or incomplete.
• It's Safe: It has built-in rules to prevent impossible medical errors.

In short: This tool takes the boring, time-consuming job of measuring heart movies and turns it into a fast, reliable, and automatic process. It frees up doctors to spend less time tracing lines and more time treating patients. And the best part? It's available for everyone to use right now.

Technical Summary

1. Problem Statement

Cardiac Magnetic Resonance (CMR) imaging is the gold standard for assessing cardiac structure and function. However, the manual segmentation of short-axis cine images to derive clinical parameters (e.g., ventricular volumes, ejection fractions, myocardial mass) is a significant bottleneck. It is time-consuming (15–45 minutes per exam), subjective, and prone to inter-observer variability.

Existing automated deep learning solutions face three critical limitations hindering clinical adoption:

1. Limited Data Scope: Most models are trained on small, single-center datasets with narrow disease coverage, leading to poor generalization on rare pathologies or severe morphological variations (e.g., dilated cardiomyopathy).
2. Restrictive Input Requirements: Many methods require 3D volumes, temporal sequences, or prior Region-of-Interest (ROI) localization, making them fragile when faced with incomplete data, varying protocols, or heterogeneous storage formats in real-world clinical settings.
3. Lack of Anatomical Plausibility: Deep learning models are statistical learners that often produce topologically impossible outputs (e.g., scattered fragments, left ventricular cavity breaching the myocardium, or gaps in the interventricular septum), eroding clinical trust.

2. Methodology

The authors developed CorSeg-CineSAX, a framework designed to maximize robustness, generalizability, and clinical usability through four key stages:

A. Dataset Construction

• Scale: The authors constructed the largest annotated CMR short-axis dataset to date, comprising 1,555 subjects from 12 centers in China.
• Diversity: The data covers five cardiac disease types (Normal, HCM, DCM, HHD, Cardiac Amyloidosis) and includes full cardiac cycle annotations (typically 25 frames per slice).
• Volume: The dataset contains 319,175 labeled 2D images.
• Annotation: Performed by experienced radiologists and reviewed by senior cardiologists, labeling the Left Ventricular Myocardium (LV Myo), Left Ventricular Cavity (LV Cav), and Right Ventricular Cavity (RV).

B. Model Architecture & Training

• Architecture: Utilized MedNeXt-L, a modern convolutional neural network inspired by ConvNeXt, featuring large-kernel convolutions and Layer Normalization.
• Strategy: A 2D slice-by-slice approach was adopted.
  • Input: Full field-of-view (no ROI cropping or localization required).
  • Independence: Processes each slice independently, eliminating dependencies on 3D context, temporal sequences, or data completeness.
• Training: Trained using a composite loss function (Dice + Focal + Cross-Entropy) with class weighting to handle background dominance. The model was trained for 1,000 epochs on NVIDIA A100 GPUs.

C. Anatomical Constraint Post-Processing

To address topological errors, a deterministic, rule-based three-step pipeline was applied after inference:

1. Connected Component Constraint: Retains only the largest connected component for each structure, removing isolated fragments.
2. Containment Relationship Constraint: Iteratively reclassifies LV Cavity pixels that directly touch the background or RV (without intervening myocardium) as LV Myocardium, ensuring the cavity is fully enclosed.
3. Gap Filling Constraint: Fills background holes completely enclosed by cardiac structures (specifically in the interventricular septum) and fills single-pixel gaps between LV Myo and RV.

D. Validation

The model was validated on:

• Internal Test Set: 310 subjects from the private dataset.
• External Test Sets: Three independent public datasets (ACDC, M&Ms1, M&Ms2) totaling 855 subjects from 6 centers across 3 countries (France, Germany, Spain) and 3 scanner vendors. These sets included 10 disease categories never seen during training.

3. Key Contributions

1. Unprecedented Data Scale: Creation of a massive, multi-center, multi-disease, full-cycle annotated dataset (319k+ images), enabling robust training for diverse cardiac morphologies.
2. Universal Input Strategy: A 2D, full-field-of-view approach that removes the need for complex preprocessing (localization, 3D stacking), ensuring compatibility with any clinical CMR acquisition protocol.
3. Anatomical Safety Net: A novel, zero-parameter post-processing pipeline that enforces anatomical priors, systematically eliminating topological violations without retraining.
4. Open-Source Deployment: Full release of source code, pre-trained weights, and deployment tools to facilitate clinical adoption and reproducibility.

4. Results

Segmentation Accuracy

• Internal Test: Mean Dice Similarity Coefficient (DSC) of 0.913 ± 0.037.
• External Test: Mean DSC of 0.911 ± 0.040.
• Generalization: The cross-domain performance gap was negligible (ΔDSC = 0.002), demonstrating robustness across different scanners, vendors, and patient populations.
• Zero-Shot Performance: On 10 unseen disease categories (e.g., Tetralogy of Fallot, Arrhythmogenic Cardiomyopathy), the model achieved DSC values between 0.899 and 0.921.

Impact of Post-Processing

While the post-processing had a negligible effect on mean DSC, it drastically improved anatomical correctness:

• Containment Violations: Reduced from 7.5% to 0%.
• Gap Errors: Reduced from 1.8% to 0%.
• Fragment Rate: Reduced by 85.5% (from 9.0% to 1.3%).
• Boundary Accuracy: The 95th percentile Hausdorff Distance (HD95) standard deviation decreased significantly (e.g., endocardial HD95 SD dropped from 8.37mm to 1.32mm), indicating the removal of severe outlier errors.

Clinical Parameter Agreement

Automated measurements showed excellent agreement with manual ground truth:

• Left Ventricle: ICC ≥ 0.982 for LVEDV, LVESV, LVEF, and LV Mass.
• Right Ventricle Volumes: ICC ≥ 0.977 for RVEDV and RVESV.
• Limitation: RVEF showed lower agreement (ICC = 0.766) due to the use of LV-based phase identification, which may not align with RV timing in cases of dyssynchrony.

5. Significance

CorSeg-CineSAX represents a significant step toward the clinical deployment of AI in cardiology. By prioritizing data scale and diversity over architectural novelty, the authors demonstrated that a proven architecture trained on a massive, heterogeneous dataset can outperform or match specialized models on benchmark challenges without needing domain-specific fine-tuning.

The framework's robustness to incomplete data (2D slice-by-slice) and anatomical guarantees (post-processing) directly address the primary barriers to clinical trust. The open-source nature of the project allows for community-driven improvement and immediate integration into clinical workflows for high-throughput phenotyping and automated reporting of cardiac function.