pith. sign in

arxiv: 2604.22258 · v1 · submitted 2026-04-24 · 💻 cs.LG · cs.AI

Protect the Brain When Treating the Heart: A Convolutional Neural Network for Detecting Emboli

Andrea Angino , Ken Trotti , Diego Ulisse Pizzagalli , Rolf Krause , Tiziano Torre , Stefanos Demertzis This is my paper

Pith reviewed 2026-05-08 12:31 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords gaseous microembolicardiac ultrasoundemboli detection2.5D U-Netconvolutional neural networkreal-time segmentationheart surgery monitoringbrain protection
0
0 comments X

The pith

A 2.5D U-Net segments gaseous microemboli in cardiac ultrasound videos with robust accuracy and real-time speed.

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

The paper develops an automated way to spot gaseous microemboli, small air bubbles that can reach the brain during heart surgery or catheter procedures. These bubbles appear in ultrasound images but are difficult to track manually because of rapid motion, operator-chosen views, and background structures that look similar. The authors apply a 2.5D U-Net to process sequences of ultrasound frames together in space and time, producing outlines of the bubbles that hold up against noise. This setup keeps processing fast enough for live use and adds the ability to measure total bubble area over time. If the method works as described, clinicians gain an objective running count that can be added directly to existing surgical monitoring systems.

Core claim

A 2.5D U-Net architecture applied to space-time connected ultrasound data yields robust detection of gaseous microemboli against background structures, high segmentation accuracy, and real-time execution speed, enabling integration into patient-monitoring surgical protocols that quantify GME area over time.

What carries the argument

The 2.5D U-Net, a convolutional network that processes short sequences of ultrasound frames to outline gaseous microemboli across both spatial position and time.

If this is right

  • The pipeline can be added to existing surgical monitoring systems without changing workflow.
  • GME area can be quantified continuously over the course of a procedure.
  • Detection stays reliable even when background structures resemble bubbles.
  • The same approach applies to both open-heart surgery and transcatheter interventions.
  • Real-time performance is preserved while segmentation accuracy remains high.

Where Pith is reading between the lines

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

  • Automated GME tracking could support future studies that link bubble volume directly to post-operative neurological outcomes.
  • The method might be retrained on other ultrasound views or for different transient objects such as solid emboli.
  • Hospitals could add threshold alerts that notify the team when cumulative GME area exceeds a chosen safety limit.
  • Standardized AI monitoring might reduce differences in emboli reporting across different surgical centers.

Load-bearing premise

The 2.5D U-Net maintains robust detection and high accuracy when applied to real clinical ultrasound data that include varying operator views and high velocities.

What would settle it

A set of live operating-room ultrasound recordings from multiple patients and operators in which the network's bubble outlines match expert manual review less than 80 percent of the time or cannot keep up with video frame rate on standard hospital hardware.

Figures

Figures reproduced from arXiv: 2604.22258 by Andrea Angino, Diego Ulisse Pizzagalli, Ken Trotti, Rolf Krause, Stefanos Demertzis, Tiziano Torre.

Figure 1
Figure 1. Figure 1: Color-coded overlay of two consecutive frames: the initial frame is depicted in red, while the subsequent frame appears in turquoise. The dynamic movement of GMEs is emphasized by their distinct coloring, contrasting with the white stationary ele￾ments representing the stable cardiac background. at echocardiographic evaluation. This similarity cre￾ates ambiguity, especially near edges, where motion is the … view at source ↗
Figure 2
Figure 2. Figure 2: Preliminary segmentation comparisons using classical Fiji-based methods and deep learning models, view at source ↗
Figure 3
Figure 3. Figure 3: Training loss over 50 epochs for 2.5D U view at source ↗
Figure 4
Figure 4. Figure 4: Segmentation performance (IoU, Dice, and FBI with standard deviation) and inference time per batch view at source ↗
Figure 5
Figure 5. Figure 5: Segmentation performance (IoU, Dice, and FBI with standard deviation) and inference time per batch view at source ↗
Figure 6
Figure 6. Figure 6: Black-and-white frame acquired in real time view at source ↗
read the original abstract

Gaseous microemboli (GME) represent a common complication of cardiac structural interventions across both surgical and transcatheter approaches. Transthoracic cardiac ultrasound imaging represents a convenient methodology to visualize the presence of circulating GME. However, their detection and quantification are far from trivial due to operator-dependent view, high velocity, and objects with similar structure in the background. Here, we propose an approach based on a 2.5D U-Net architecture to segment GME in space-time connected data. Such an approach yields robust detection against the background and high segmentation accuracy while retaining real-time execution speed. These properties facilitated the integration of the proposed pipeline into patient-monitoring surgical protocols, providing the quantification of GME area over 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 / 2 minor

Summary. The manuscript proposes a 2.5D U-Net architecture to segment gaseous microemboli (GME) in space-time connected transthoracic ultrasound data. It claims this yields robust background rejection, high segmentation accuracy (via Dice/IoU), and real-time inference (sub-30 ms) sufficient for integration into surgical monitoring protocols to quantify GME area over time.

Significance. If the reported metrics generalize, the work offers a practical tool for real-time emboli monitoring during cardiac interventions, addressing a clinically relevant complication. The temporal stacking in the 2.5D design and timing benchmarks are strengths that support feasibility; however, the overall significance is moderated by the absence of external validation or baselines, limiting claims of robustness across varying clinical conditions.

major comments (2)
  1. [Results] Results section: Performance (Dice/IoU) is reported only on an internal train-test split of the dataset. This does not address the operator-dependent views, high velocities, and background objects emphasized in the Introduction as core challenges; without patient-wise cross-validation, multi-center data, or external test sets, the central claim of 'robust detection' lacks sufficient support.
  2. [Methods] Methods section: The 2.5D U-Net with temporal stacking is described as standard, but no comparisons to 2D U-Net, 3D U-Net, or non-DL baselines (e.g., intensity thresholding) are provided. This makes it impossible to quantify the specific benefit of the proposed architecture for the space-time segmentation task.
minor comments (2)
  1. [Abstract] Abstract: Claims of 'high segmentation accuracy' and 'robust detection' are made without any numerical results (e.g., Dice score or timing); adding these would immediately ground the assertions.
  2. [Methods] Figure captions and Methods: Clarify the exact temporal stacking (number of frames) and augmentation strategy used for the space-time volumes; this detail is needed to reproduce the real-time performance.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We are grateful to the referee for their constructive comments on our manuscript. We have revised the paper to address the concerns raised and provide point-by-point responses below.

read point-by-point responses

  1. Referee: [Results] Results section: Performance (Dice/IoU) is reported only on an internal train-test split of the dataset. This does not address the operator-dependent views, high velocities, and background objects emphasized in the Introduction as core challenges; without patient-wise cross-validation, multi-center data, or external test sets, the central claim of 'robust detection' lacks sufficient support.

    Authors: We acknowledge the importance of robust evaluation for the claimed detection performance. In the revised manuscript, we have incorporated patient-wise cross-validation results to better account for variability across patients, operators, and views. These additional experiments, reported in the updated Results section, demonstrate consistent performance metrics, thereby supporting the robustness against the challenges outlined in the Introduction. We note that multi-center or external test sets are not available for this study and have explicitly discussed this as a limitation in the revised Discussion section. revision: partial

  2. Referee: [Methods] Methods section: The 2.5D U-Net with temporal stacking is described as standard, but no comparisons to 2D U-Net, 3D U-Net, or non-DL baselines (e.g., intensity thresholding) are provided. This makes it impossible to quantify the specific benefit of the proposed architecture for the space-time segmentation task.

    Authors: We agree that comparative analyses are necessary to highlight the benefits of the 2.5D architecture. Accordingly, we have added comparisons against 2D U-Net, 3D U-Net, and an intensity-based thresholding baseline in the revised Methods and Results sections. The quantitative results show that the temporal component in the 2.5D U-Net provides improved segmentation accuracy for high-velocity GME while maintaining real-time performance, as now detailed in the manuscript. revision: yes

standing simulated objections not resolved
  • The lack of multi-center or external validation datasets, which cannot be addressed without collecting new data from additional centers.

Circularity Check

0 steps flagged

No circularity; empirical application of standard 2.5D U-Net

full rationale

The manuscript applies a 2.5D U-Net to space-time ultrasound data for GME segmentation. No derivation chain, uniqueness theorem, or parameter-fitting step is present that reduces a claimed prediction or result to its own inputs by construction. Methods consist of standard architecture, loss, augmentation, and internal-split evaluation; reported Dice/IoU and timing figures are direct empirical outcomes rather than self-referential. No self-citations are load-bearing for any mathematical claim, and no ansatz or renaming of known results occurs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Only abstract available; typical assumptions for ML medical imaging papers apply, such as the suitability of CNNs for image segmentation tasks.

axioms (1)
  • domain assumption Ultrasound images can be treated as space-time connected data for segmentation.
    The paper relies on this to use 2.5D architecture.

pith-pipeline@v0.9.0 · 5438 in / 1175 out tokens · 59505 ms · 2026-05-08T12:31:52.401172+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

18 extracted references · 1 canonical work pages

  1. [1]

    Projection- based 2.5d u-net architecture for fast volumetric segmenta- tion.CoRR, abs/1902.00347, 2019

    Christoph Angermann, Markus Haltmeier, Ruth Steiger, Sergiy Pereverzyev Jr., and Elke Ruth Gizewski. Projection- based 2.5d u-net architecture for fast volumetric segmenta- tion.CoRR, abs/1902.00347, 2019

    work page arXiv 1902

  2. [2]

    Angino and K

    A. Angino and K. Trotti. Aircatch. https://github.com/ AnginoA/Aircatch, 2025. GitHub repository

    2025

  3. [3]

    Paola Antonello, Diego Morone, Edisa Pirani, Mariagrazia Uguccioni, Marcus Thelen, Rolf Krause, and Diego Piz- zagalli. Tracking unlabeled cancer cells imaged with low resolution in wide migration chambers via u-net class-1 probability (pseudofluorescence).Journal of Biological En- gineering, 17, 01 2023

    2023

  4. [4]

    Picie: Unsupervised semantic segmentation using invariance and equivariance in clustering

    Jang Hyun Cho, Utkarsh Mall, Kavita Bala, and Bharath Hariharan. Picie: Unsupervised semantic segmentation using invariance and equivariance in clustering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16794–16804, June 2021

    2021

  5. [5]

    Emma M. L. Chung, Caroline Banahan, Nikil Patel, Justyna Janus, David Marshall, Mark A. Horsfield, Cl ´ement Rousseau, Jonathan Keelan, David H. Evans, and James P. Hague. Size distribution of air bubbles entering the brain during cardiac surgery.PLOS ONE, 10(4):1–11, 04 2015

    2015

  6. [6]

    Dueholm Vestergaard, U

    C. Dueholm Vestergaard, U. Elstrøm Vindelev, L. P. Muren, J. Ren, O. Nørrevang, K. Jensen, and V . T. Taasti. Data augmentation for medical imaging: A systematic literature re- view.Physics Imaging and Radiation Oncology, 32:100658, 2024

    2024

  7. [7]

    Pylv ¨an¨ainen, St´ephane U

    Dmitry Ershov, Minh-Son Phan, Joanna W. Pylv ¨an¨ainen, St´ephane U. Rigaud, Laure Le Blanc, Arthur Charles-Orszag, James R. W. Conway, Romain F. Laine, Nathan H. Roy, Daria Bonazzi, Guillaume Dum ´enil, Guillaume Jacquemet, and Jean-Yves Tinevez. Trackmate 7: integrating state-of-the- art segmentation algorithms into tracking pipelines.Nature Methods, 19...

    2022

  8. [8]

    Mask r-cnn

    Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross Gir- shick. Mask r-cnn. InProceedings of the IEEE International Conference on Computer Vision (ICCV), October 2017

    2017

  9. [9]

    Henriques, and Andrea Vedaldi

    Xu Ji, Joao F. Henriques, and Andrea Vedaldi. Invariant infor- mation clustering for unsupervised image classification and segmentation. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019

    2019

  10. [10]

    Transformers in medical image segmentation: a narrative review.Quantitative Imaging in Medicine and Surgery, 13(12), 2023

    Rabeea Fatma Khan, Byoung-Dai Lee, and Mu Sook Lee. Transformers in medical image segmentation: a narrative review.Quantitative Imaging in Medicine and Surgery, 13(12), 2023

    2023

  11. [11]

    Investigation of air bubble properties: Relevance to prevention of coronary air embolism during cardiac surgery.Artificial Organs, 45(9):E349–E358, September 2021

    Kazuki Kihara and Kazumasa Orihashi. Investigation of air bubble properties: Relevance to prevention of coronary air embolism during cardiac surgery.Artificial Organs, 45(9):E349–E358, September 2021. © 2021 International Center for Artificial Organs and Transplantation and Wiley Periodicals, Inc

    2021

  12. [12]

    A. Kurz, K. Hauser, H. A. Mehrtens, E. Krieghoff-Henning, A. Hekler, J. N. Kather, S. Fr ¨ohling, C. von Kalle, and T. J. Brinker. Uncertainty estimation in medical image classification: Systematic review.JMIR Medical Informatics, 10(8):e36427, 2022

    2022

  13. [13]

    Self- supervised vision transformers are efficient segmentation learners for imperfect labels, 2024

    Seungho Lee, Seoungyoon Kang, and Hyunjung Shim. Self- supervised vision transformers are efficient segmentation learners for imperfect labels, 2024

    2024

  14. [14]

    Fully convolutional networks for semantic segmentation

    Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. InPro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2015

    2015

  15. [15]

    An analytics-driven review of u-net for medical image seg- mentation.Healthcare Analytics, 8:100416, 2025

    Fnu Neha, Deepshikha Bhati, Deepak Kumar Shukla, Son- avi Makarand Dalvi, Nikolaos Mantzou, and Safa Shubbar. An analytics-driven review of u-net for medical image seg- mentation.Healthcare Analytics, 8:100416, 2025

    2025

  16. [16]

    Retained intrac- ardiac air in cardiovascular surgery: a re-visited problem

    Kazumasa Orihashi and Tsuyoshi Miyata. Retained intrac- ardiac air in cardiovascular surgery: a re-visited problem. General Thoracic and Cardiovascular Surgery, 72(7):429– 438, July 2024. Published online 2024/07/01

    2024

  17. [17]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Nassir Navab, Joachim Hornegger, William M. Wells, and Alejandro F. Frangi, editors,Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, pages 234– 241, Cham, 2015. Springer International Publishing

    2015

  18. [18]

    Elkin, and Vijay Devabhaktuni

    Nahian Siddique, Sidike Paheding, Colin P. Elkin, and Vijay Devabhaktuni. U-net and its variants for medical image segmentation: A review of theory and applications.IEEE Access, 9:82031–82057, 2021

    2021