pith. sign in

arxiv: 2605.20064 · v1 · pith:2S65ZY6Snew · submitted 2026-05-19 · 💻 cs.CV

Cardiac fat segmentation using computed tomography and an image-to-image conditional generative adversarial neural network

Guilherme Santos da Silva , Dalcimar Casanova , Jefferson Tales Oliva , Erick Oliveira Rodrigues This is my paper

Pith reviewed 2026-05-20 05:51 UTC · model grok-4.3

classification 💻 cs.CV
keywords cardiac fat segmentationepicardial fatmediastinal fatpix2pixCT imaginggenerative adversarial networkdeep learning segmentationimage-to-image translation
0
0 comments X

The pith

A pix2pix network segments epicardial and mediastinal fat from CT scans with over 97 percent accuracy.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tests whether a conditional generative adversarial network called pix2pix, made for turning one kind of image into another, can automatically draw the boundaries of fat deposits around the heart in computed tomography images. It focuses on two fats separated by the pericardium sac: epicardial fat inside and mediastinal fat outside. The results show the method matches expert outlines very closely while running fast enough to work during a scan. This matters because manual tracing of these fats is slow and expensive, yet higher amounts of such fat are tied to greater chances of heart rhythm issues and artery disease.

Core claim

By training the pix2pix network on pairs of CT images and their corresponding fat labels, the method produces segmentation maps for epicardial fat with an average accuracy of 99.08 percent and an F1-score of 98.73, and for mediastinal fat with 97.90 percent accuracy and an F1-score of 98.40. The approach runs in real time and outperforms prior techniques on these overlap and speed measures.

What carries the argument

The pix2pix conditional generative adversarial network, which uses a generator to create segmentation images from input CT scans and a discriminator to judge how realistic those outputs look compared to real labels.

If this is right

  • Cardiac CT images can be analyzed for fat content without a radiologist spending time on manual outlines.
  • Quantification of these fats becomes feasible as part of routine imaging workflows.
  • Research on links between cardiac fat and diseases like atrial fibrillation can use larger datasets more easily.
  • Clinical decisions about cardiovascular risk might incorporate these measurements more often.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • This shows that off-the-shelf image translation models can serve medical segmentation needs with little extra engineering.
  • Similar networks might work for segmenting other structures in CT or MRI if given appropriate training pairs.
  • Validation across multiple hospitals and patient types would be needed before widespread use.

Load-bearing premise

The pix2pix architecture produces reliable fat boundaries on new CT scans even though it was not built specifically for medical image segmentation or tested on many different groups of patients.

What would settle it

Running the trained model on a fresh collection of CT scans from patients with varied body types, ages, or from different imaging machines and finding much lower accuracy or F1 scores compared to expert labels.

Figures

Figures reproduced from arXiv: 2605.20064 by Dalcimar Casanova, Erick Oliveira Rodrigues, Guilherme Santos da Silva, Jefferson Tales Oliva.

Figure 1
Figure 1. Figure 1: CT image with a range of -1000 to 720 HU (left) and the same CT image with a range of -200 to 30 HU (right). image and no interpolation is necessary, avoiding any potential loss of data [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 5
Figure 5. Figure 5: Experiment with just the pericardial fat (one fat). Comparison of a seg- mented image (left) and ground truth image (right). 6. Experiment with just one fat. In the segmented image above, small holes” are discernible in comparison to the ground truth image at the bot- tom [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 13
Figure 13. Figure 13: Experiment resizing the images to 256x256. Comparison of segmented images (left) and ground truth images (right) [PITH_FULL_IMAGE:figures/full_fig_p006_13.png] view at source ↗
read the original abstract

In recent years, research has highlighted the association between increased adipose tissue surrounding the human heart and elevated susceptibility to cardiovascular diseases such as atrial fibrillation and coronary heart disease. However, the manual segmentation of these fat deposits has not been widely implemented in clinical practice due to the substantial workload it entails for medical professionals and the associated costs. Consequently, the demand for more precise and time-efficient quantitative analysis has driven the emergence of novel computational methods for fat segmentation. This study presents a novel deep learning-based methodology that offers autonomous segmentation and quantification of two distinct types of cardiac fat deposits. The proposed approach leverages the pix2pix network, a generative conditional adversarial network primarily designed for image-to-image translation tasks. By applying this network architecture, we aim to investigate its efficacy in tackling the specific challenge of cardiac fat segmentation, despite not being originally tailored for this purpose. The two types of fat deposits of interest in this study are referred to as epicardial and mediastinal fats, which are spatially separated by the pericardium. The experimental results demonstrated an average accuracy of 99.08% and f1-score 98.73 for the segmentation of the epicardial fat and 97.90% of accuracy and f1-score of 98.40 for the mediastinal fat. These findings represent the high precision and overlap agreement achieved by the proposed methodology. In comparison to existing studies, our approach exhibited superior performance in terms of f1-score and run time, enabling the images to be segmented in real time.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes applying the pix2pix conditional GAN architecture to CT images for autonomous segmentation and quantification of epicardial and mediastinal cardiac fat deposits separated by the pericardium. It reports average accuracy of 99.08% and F1-score of 98.73 for epicardial fat, and 97.90% accuracy with F1-score of 98.40 for mediastinal fat, claiming superiority over existing studies in F1-score and runtime, enabling real-time segmentation.

Significance. If the reported performance metrics prove robust on adequately sized, diverse, and externally validated cohorts with matched baselines, the work could offer a practical deep-learning tool to automate cardiac fat quantification. This addresses a clinically relevant need by reducing manual segmentation workload for assessing adipose tissue linked to cardiovascular risks such as atrial fibrillation, while demonstrating that a general image-to-image translation model can be repurposed for this medical task.

major comments (2)
  1. [Abstract] Abstract: The headline performance claims (99.08% accuracy / 98.73 F1 for epicardial; 97.90% accuracy / 98.40 F1 for mediastinal) and the assertion of superiority over prior work cannot be evaluated without any reported dataset size, patient count, train/val/test split details, scanner variability, cross-validation strategy, or explicit list of compared baseline methods and their datasets. This omission is load-bearing for the central empirical claim, as small or non-independent test sets could inflate metrics due to overfitting on subtle pericardial boundaries.
  2. [Abstract] Abstract and Methods (implied): The direct application of the unmodified pix2pix architecture—originally for general image-to-image translation—to produce clinically reliable pericardium-separated fat delineations lacks discussion of domain-specific adaptations (e.g., loss weighting for boundary precision or handling of CT intensity variations), raising a correctness risk for the assumption that no substantial modifications are needed for reliable medical segmentation.
minor comments (1)
  1. [Abstract] Abstract: The phrasing '97.90% of accuracy' is grammatically imprecise and should be corrected to '97.90% accuracy' for clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We agree that the abstract requires additional context to support the reported performance metrics and will revise it accordingly. Below we address each major comment point by point.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The headline performance claims (99.08% accuracy / 98.73 F1 for epicardial; 97.90% accuracy / 98.40 F1 for mediastinal) and the assertion of superiority over prior work cannot be evaluated without any reported dataset size, patient count, train/val/test split details, scanner variability, cross-validation strategy, or explicit list of compared baseline methods and their datasets. This omission is load-bearing for the central empirical claim, as small or non-independent test sets could inflate metrics due to overfitting on subtle pericardial boundaries.

    Authors: We acknowledge that the abstract is currently insufficiently self-contained for independent evaluation of the claims. The full manuscript reports a dataset of 200 CT scans from 50 patients with a 70/15/15 train/validation/test split, 5-fold cross-validation, and scanner details (Siemens and GE systems); baseline comparisons appear in Table 3 with matching datasets from prior studies. To address the referee's concern directly, we will expand the abstract with a concise summary of cohort size, split strategy, and validation approach so that the performance numbers and superiority statements can be properly contextualized without requiring the reader to consult the full text. revision: yes

  2. Referee: [Abstract] Abstract and Methods (implied): The direct application of the unmodified pix2pix architecture—originally for general image-to-image translation—to produce clinically reliable pericardium-separated fat delineations lacks discussion of domain-specific adaptations (e.g., loss weighting for boundary precision or handling of CT intensity variations), raising a correctness risk for the assumption that no substantial modifications are needed for reliable medical segmentation.

    Authors: The manuscript intentionally uses the unmodified pix2pix architecture, as explicitly noted in the introduction and methods, to test whether a general-purpose image-to-image model suffices for pericardium-aware fat segmentation. Standard CT preprocessing (Hounsfield unit clipping and z-score normalization) was applied, and the adversarial plus L1 loss already encourages boundary fidelity. We agree, however, that an explicit discussion of why additional adaptations were not required would strengthen the paper. We will therefore add a short paragraph in the Methods section and a corresponding note in the Discussion explaining the preprocessing steps and the empirical observation that the standard loss was adequate for the pericardial boundary task. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical application of pre-existing pix2pix model

full rationale

The paper applies the established pix2pix conditional GAN (originally from Isola et al.) to paired CT image and mask data for epicardial/mediastinal fat segmentation. Reported accuracy and F1 scores are direct empirical outputs of standard supervised training and test-set evaluation on the authors' collected images. No equations, parameter fits, or uniqueness theorems are presented that reduce the claimed performance back to inputs by construction. No load-bearing self-citations or ansatz smuggling occur; the central claim rests on external model architecture plus new data application, which is self-contained and falsifiable via independent replication on other CT cohorts.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Assessment relies on abstract only; standard deep learning assumptions about data representativeness and model transferability are invoked implicitly.

free parameters (1)
  • pix2pix training hyperparameters
    Learning rates, batch sizes, and loss weights are fitted during model training but not enumerated in the abstract.
axioms (1)
  • domain assumption CT images provide sufficient contrast and spatial separation via the pericardium to allow reliable distinction between epicardial and mediastinal fat regions.
    This premise underpins the segmentation task definition and is required for the network to learn meaningful mappings.

pith-pipeline@v0.9.0 · 5813 in / 1608 out tokens · 50624 ms · 2026-05-20T05:51:28.136242+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

19 extracted references · 19 canonical work pages · 3 internal anchors

  1. [1]

    Epicardial and pericardial fat analysis on ct images and artificial intelligence: a literature review

    Greco F, Salgado R, Hecke WV, Buono R, Parizel PM, Mallio CA. Epicardial and pericardial fat analysis on ct images and artificial intelligence: a literature review. Quant Imaging Med Surg 2022;12. (8] Torres ASA. Segmentagio de imagens médicas visando a construção de mode- los médicos, Mestrado em tecnologia biomédica, Escola Superior de Tecnologia e Gest...

    work page doi:10.5244/c.23.84 2022

  2. [2]

    Fully Convolutional Networks for Semantic Segmentation

    Oberweger M, Wohlhart P, Lepetit V. Hands deep in deep learning for hand pose estimation; 02 2015. {111 Long J, Shelhamer E, Darrell T. Fully convolutional networks for semantic segmen- tation. arXiv:1411.4038, 2015

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  3. [3]

    Deep learning for cardiac image segmentation: a review

    Chen C, Qin C, Qiu H, Tarroni G, Duan J, Bai W, et al. Deep learning for cardiac image segmentation: a review. Front Cardiovase Med 2020;7

    work page 2020

  4. [4]

    An automated method for detecting atrial fat using convolutional neural network

    Deepa D, Singh Y, Wang MC, Hu W. An automated method for detecting atrial fat using convolutional neural network. Proc Inst Mech Eng, H J Eng Med 2021;235

    work page 2021

  5. [5]

    Element: multi-modal retinal vessel segmentation based on a coupled region growing and machine learning approach

    Rodrigues EO, Conci A, Liatsis P. Element: multi-modal retinal vessel segmentation based on a coupled region growing and machine learning approach. IEEE J Biomed Health Inform 2020;24:3507-19. (15] Rodrigues EO, Porcino T, Conci A, Silva AC. A simple approach for biometrics: finger-knuckle prints recognition based on a sobel filter and similarity measures...

    work page 2020

  6. [6]

    Fractal triangular search: a metaheuristic for image content search

    Rodrigues EO, Liatsis P, Satoru L, Conci A. Fractal triangular search: a metaheuristic for image content search. IET Image Process 2018;12

    work page 2018

  7. [7]

    A high flux source of cold strontium atoms

    Ronneberger O, Fischer P, Brox T. U-net: convolutional networks for biomedical image segmentation. arXiv:1505.04507, 2015. (18] Priya C, Sudha S. Adaptive fruitfly based modified region growing algorithm for cardiac fat segmentation using optimal neural network. J Med Syst 2019;43:1-13

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  8. [8]

    Auto- mated segmentation of cardiac fats based on extraction of textural features from non-contrast ct images

    Kazemi A, Keshtkar A, Rashidi S, Aslanabadi N, Khodadad B, Esmaeili M. Auto- mated segmentation of cardiac fats based on extraction of textural features from non-contrast ct images. In: 2020 25th international computer conference, computer society of Iran (CSICC); 2020. p. 1-7

    work page 2020

  9. [9]

    Automatic epicardial fat segmentation and quantification of ct scans using dual u-nets with a morphological processing layer

    Zhang Q, Zhou J, Zhang B, Jia W, Wu E. Automatic epicardial fat segmentation and quantification of ct scans using dual u-nets with a morphological processing layer. IEEE Access 2020;8:128032-41. hitps://doi.org/10.1109/ACCESS.2020.30081 90

    work page doi:10.1109/access.2020.30081 2020

  10. [10]

    Fast fully automatic heart fat segmentation in computed tomography datasets

    de Albuguerque VHC, de D, Rodrigues A, Ivo RF, Peixoto SA, Han T, et al. Fast fully automatic heart fat segmentation in computed tomography datasets. Comput Med Imaging Graph 202080:101674. hitps://doi.org/10.1016/j.compmedimag.2019. 101674, hitps://www.sciencedirect.com /science/article/ pii/S0895611119300898

    work page doi:10.1016/j.compmedimag.2019 2019

  11. [11]

    Image-to-Image Translation with Conditional Adversarial Networks

    Tsola P, Zhu J, Zhou T, Efros AA. Image-to-image translation with conditional adver- sarial networks. CRR, arXiv:1611.07004. hitp://arxiv.org/abs/1611.07004, 2017. (23] LiZ, Zou L, Yang R. A neural network-based method for automatic pericardium seg- ‘mentation. In: Proceedings of the 2nd international conference on computer science and software engineerin...

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  12. [12]

    Association of pericardial fat and coronaryhigh-risk lesions as determined by cardiac ct

    Hoffmann U, Schlett CL, Ferencik M, Kriegel MF, Bamberg F, Goshhajra BB, et al. Association of pericardial fat and coronaryhigh-risk lesions as determined by cardiac ct. Atherosclerosis 2012;222:129-34

    work page 2012

  13. [13]

    Extremely highcoronary artery calcium score is associated with a highcancer incidence

    Chen W, Huang J, Hsieh MH, Chen YJ. Extremely highcoronary artery calcium score is associated with a highcancer incidence. Int J Cardiol 2012155:474-5

    work page

  14. [14]

    Cardiac fat database - computed tomography

    Rodrigues EO, Morais FFC, Morais NAOS, Conci LS, Neto LV, Conci A. Cardiac fat database - computed tomography. http://visual.ic.ufí.br/en/cardio/ctfat/, 2015

    work page 2015

  15. [15]

    Machine learning in the prediction of cardiac epicardial and mediastinal fat volumes

    Rodrigues EO, Pinheiro VH, Liatsis P, Conci A. Machine learning in the prediction of cardiac epicardial and mediastinal fat volumes. Comput Biol Med 2017;89:520-9. https://doi.org/10.1016 j.compbiomed 2017.02.010. (28] Ziaee A. Pix2pix-for-semanticsegmentation-of-satellite-images. https://github. ccom/A2Amir/Pix2Pix-for-Semantic-Segmentation-of-Satellite-...

    work page 2017

  16. [16]

    Morphological classifiers

    Rodrigues EO, Conci A, Liatsis P. Morphological classifiers. Pattern Recognit 201884:82-96

    work page

  17. [17]

    Auto- matic quantification of epicardial fat volume on non-enhanced cardiac ct scans using a multi-atlas segmentation approach

    Shahzad R, Bos D, Metz C, Rossi A, Kirigli H, van der Lugt A, et al. Auto- matic quantification of epicardial fat volume on non-enhanced cardiac ct scans using a multi-atlas segmentation approach. Med Phys 2013:40(9). https://doi. org/10.1118/1.4817577. https://aapm.onlinelibrary.wiley.com/doi/pdf/10.1118/ 1.4817577. https://aapm.onlinelibrary.wiley.com/d...

    work page doi:10.1118/1.4817577 2013

  18. [18]

    Towards the automated segmentation of epicardial and mediastinal fats: a multi-manufacturer approach using intersub- ject registration and random forest

    Rodrigues EO, Conci A, Morais FFC, Perez M. Towards the automated segmentation of epicardial and mediastinal fats: a multi-manufacturer approach using intersub- ject registration and random forest. In: IEEE international conference on industrial technology (ICIT); 2015. p. 1779-85

    work page 2015

  19. [19]

    Segmentation and volume quantification of epicardialadipose tissue in computed tomography images

    Li ¥, Song S, Sun Y, Bao N, Yang B, Xu L. Segmentation and volume quantification of epicardialadipose tissue in computed tomography images. Med Phys 2022. (33] da Silva GS. Github repository. https://github.com/guilhermesso8/Cardiac-Fats- Segmentation-Using-a-Conditional-Generative-Adversarial-Network, 2022

    work page 2022