💻 computer science

Towards the automated segmentation of epicardial and mediastinal fats: A multi-manufacturer approach using intersubject registration and random forest

This paper proposes a fully automated, multi-manufacturer method using intersubject registration and random forest classification to segment epicardial and mediastinal fats on coronary calcium scoring CT images, achieving high accuracy (98.4%) and a Dice similarity index of 96.8%.

Original authors: É. O. Rodrigues, A. Conci, F. F. C. Morais, M. G. Pérez

Published 2026-05-29

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: É. O. Rodrigues, A. Conci, F. F. C. Morais, M. G. Pérez

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 your heart isn't just a pump, but a house surrounded by two different types of "landscapes." Inside the fence of the house (the pericardium) is one type of fat, called epicardial fat. Outside that fence, but still within the neighborhood (the mediastinum), is another type, called mediastinal fat.

Doctors know that the amount of these specific fats is a big clue about heart health, often more so than how much fat a person has on the rest of their body. However, measuring them is like trying to count the grains of sand on a beach while the wind is blowing; it's messy, and if two different people do the counting, they often get different results.

This paper presents a new, fully automatic computer program designed to be a super-accurate "sand counter" for these heart fats using CT scans (3D X-ray pictures).

Here is how their "robot assistant" works, broken down into simple steps:

1. The "Group Photo" Trick (Registration)

Imagine you have a stack of photos of different people's hearts. The problem is that in some photos, the heart is tilted left; in others, it's shifted right or zoomed in. If you try to compare them directly, it's a mess.

The authors' first step is to line everyone up. They created a "standard template" (an atlas) by taking a specific landmark—the area behind the breastbone (retrosternal area)—from 10 different people and blending them into one average picture.

Then, the computer looks at a new patient's scan and tries to find that same landmark. To make sure it doesn't get confused by a weird angle, it uses a clever "heuristic" check: it draws two imaginary lines from the center of that landmark outward. If those lines hit fat in a specific, logical pattern, the computer knows, "Okay, I've found the right spot!" It then shifts and scales the whole image so the heart sits in the exact same position as everyone else's. This is like taking a group of people standing in a messy circle and gently nudging them all to stand in a perfect, straight line facing the camera.

2. The "Pixel Detective" (Classification)

Once the hearts are lined up, the computer needs to decide: "Is this specific pixel of fat inside the fence (epicardial) or outside the fence (mediastinal)?"

Instead of just looking at how bright or dark a pixel is (which can be tricky), the computer acts like a detective gathering clues. For every single pixel, it looks at:

Its brightness.
Its exact location.
The "texture" of its neighbors (is the fat smooth or bumpy?).

It feeds all these clues into a Random Forest algorithm. Think of a Random Forest as a council of 100 different experts. Each expert looks at the clues and votes: "This is epicardial fat!" or "This is mediastinal fat!" The computer takes a vote, and the majority wins. This method is very good at avoiding mistakes because it doesn't rely on just one rule.

3. The Results: How Good Was It?

The authors tested this system on 82 different patients (using scans from two different machine manufacturers, Siemens and Philips, to make sure it works generally).

Success Rate: The computer found the correct spot to start its analysis 100% of the time.
Accuracy: When it came to actually coloring in the fat, it was right about 98.4% of the time.
Agreement: If you compared the computer's drawing to a drawing made by human experts, they matched up 96.8% of the time (a score known as the Dice index).

The Catch

The paper is very honest about one downside: Speed. While the computer is incredibly accurate, it is currently slow. On a standard home computer, it takes several hours to process one patient's full scan. The authors suggest that if they could figure out which clues are the most important and ignore the rest, they could make the robot much faster.

In summary: The paper introduces a computer method that automatically lines up heart scans and uses a "voting council" of algorithms to perfectly separate two types of heart fat. It is highly accurate and requires no human help to start, but it currently takes a long time to finish the job.

Technical Summary: Towards the Automated Segmentation of Epicardial and Mediastinal Fats

Problem Statement
The quantification of cardiac adipose tissues, specifically epicardial and mediastinal fats, is clinically significant as these tissues correlate with various cardiovascular risk factors, including coronary artery calcification, atrial fibrillation, and carotid stiffness. Unlike general body fat, cardiac fat varies independently of a subject's overall adiposity, necessitating specific quantitative analysis. While Computed Tomography (CT) offers superior spatial resolution for this task compared to MRI or Echocardiography, manual segmentation is highly prone to inter- and intra-observer variability. Existing automated or semi-automated methods often suffer from low success rates, require significant user intervention (e.g., placing control points), or lack robustness across different CT manufacturers. This work addresses the need for a fully automatic, reproducible method to segment both epicardial and mediastinal fats from standard coronary calcium scoring CT scans with minimal user intervention.

Methodology
The proposed approach consists of two primary stages: intersubject image registration and classification-based segmentation.

Intersubject Registration:
To standardize the positioning of hearts across different patients and manufacturers (Siemens and Philips) without manual landmark placement, the authors developed an autonomous registration method.
- Atlas Construction: An atlas of the retrosternal area (the region behind the sternum) was created by averaging 10 manually aligned retrosternal slices from distinct patients, restricted to the fat intensity range (0 HU and higher).
- Recognition: The system identifies the retrosternal area in new patient slices by sliding the atlas and computing a similarity score using Weighted Mutual Information (WMI).
- Heuristic Confirmation: To improve the initial 70% recognition rate of WMI, a heuristic confirmation method was introduced. This method verifies the position by checking for the continuity of fat depots. It defines a central rectangular area and traces two points ( $p_l$ and $p_r$ ) in the bottom-left and bottom-right directions. The position is validated if the resulting "binding" line meets specific width, displacement, and distance criteria relative to the image width. This combination raised the recognition success rate to 100%.
- Transformation: Once the retrosternal center is identified, an affine transformation is applied to translate all slices of the patient to a common centralized point, normalizing the heart's position.
Classification-Based Segmentation:
Following registration, the segmentation is performed using a supervised machine learning approach (pixel classification).
- Ground Truth: Two specialists (a physicist and a computer scientist) manually segmented the epicardial and mediastinal fats in 20 patients (1000 slices) to create a ground truth dataset.
- Feature Extraction: For each pixel, a feature vector was constructed including:
  - Pixel grey level.
  - Absolute and relative (to the center of gravity) spatial coordinates ( $x, y, z$ ).
  - Texture-based features derived from a surrounding window (e.g., 25x25 pixels), including arithmetic mean, co-occurrence matrix moments, geometrical moments, run percentage, grey level non-uniformity, and a Coefficient of Smooth Variation (CSV).
- Algorithm Selection: Various classifiers from the Weka library were evaluated. While J48Graft showed high accuracy, the Random Forest algorithm was selected as the optimal balance between accuracy and segmentation sparsity.

Key Contributions

Fully Automatic Workflow: The method eliminates the need for manual control points or landmark placement, a common bottleneck in previous semi-automated approaches.
Multi-Manufacturer Robustness: The approach was validated on CT data from two different manufacturers (Siemens and Philips), demonstrating adaptability to varying acquisition protocols.
Dual-Tissue Segmentation: Unlike many prior works focusing solely on epicardial fat, this method simultaneously segments both epicardial and mediastinal fats, distinguishing them based on the pericardium boundary.
Public Ground Truth: The authors provide a publicly available dataset of manually segmented cardiac CT images and the corresponding feature data, addressing a lack of standardized benchmarks in the field.

Results
The method was evaluated using a dataset of approximately 3 GB derived from 20 patients. Performance was assessed using 66% split and 10-fold cross-validation test modes.

Accuracy: The mean accuracy for both epicardial and mediastinal fat segmentation was 98.4%.
True Positive Rate: The mean true positive rate was 96.2%.
Dice Similarity Index: The average Dice similarity index was 96.8%.
Comparison: Compared to existing literature (e.g., Shahzad et al. with 89.15% Dice, Ding et al. with 93.0% Dice), this work achieved a higher Dice score (97.9% for epicardial fat in the 66% split) and a 100% successful automatic segmentation rate (82/82 instances), whereas other fully automatic methods reported lower success rates or did not provide this metric.

Significance and Limitations
The paper claims that the integration of intersubject registration with Random Forest classification offers a robust solution for the automated quantification of cardiac fats, potentially enhancing the prognostic value of cardiac CT trials and reducing observer variability. The authors note that the Random Forest algorithm outperformed neural networks and other function-based classifiers in their analysis.

However, the authors modestly acknowledge a significant limitation: processing time. On a standard personal computer (dual-core, 4GB RAM), the algorithm currently takes several hours to fully segment a single patient due to the extensive feature extraction and large dataset size. The authors suggest that future work should focus on extensive feature evaluation to select only the most important attributes, which would reduce the dataset size and accelerate the classification step.

1. The "Group Photo" Trick (Registration)

2. The "Pixel Detective" (Classification)

3. The Results: How Good Was It?

The Catch

More like this