pith. sign in

arxiv: 2604.07780 · v2 · submitted 2026-04-09 · 📡 eess.IV · cs.CV

MonoUNet: A Robust Tiny Neural Network for Automated Knee Cartilage Segmentation on Point-of-Care Ultrasound Devices

Alvin Kimbowa , Arjun Parmar , Ibrahim Mujtaba , Will Wei , Maziar Badii , Matthew Harkey , David Liu , Ilker Hacihaliloglu This is my paper

Pith reviewed 2026-05-12 01:46 UTC · model grok-4.3

classification 📡 eess.IV cs.CV
keywords knee cartilage segmentationpoint-of-care ultrasoundlightweight neural networkmonogenic featuresdeep learningU-Netosteoarthritisimage segmentation
0
0 comments X

The pith

MonoUNet combines a stripped-down U-Net with trainable monogenic phase features and gating to segment knee cartilage on portable ultrasound devices at high accuracy with far less computation.

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

The paper introduces MonoUNet to automate knee cartilage segmentation in ultrasound images captured on point-of-care devices. It starts with an aggressively reduced U-Net backbone, adds a trainable monogenic block to extract multi-scale local phase information from the raw images, and uses a gating mechanism to feed those features into the encoder stages. This design yields Dice scores between 92.62% and 94.82%, surface distance errors of 0.133 to 0.254 mm, and strong agreement with manual thickness and echo-intensity readings while cutting parameters by 10x to 700x and compute by 14x to 2000x relative to other lightweight models. A sympathetic reader would care because the resulting model fits on resource-limited portable hardware and could support routine osteoarthritis monitoring without large equipment or cloud processing.

Core claim

MonoUNet demonstrates that a highly compact segmentation network built from a reduced U-Net, a trainable monogenic block for multi-scale local phase extraction, and a gating mechanism that injects those features into encoder stages can achieve robust performance on multi-site, multi-device knee cartilage ultrasound data, with average Dice scores of 92.62% to 94.82%, MASD values of 0.133 mm to 0.254 mm, and excellent reliability (ICC 0.96 for thickness, 0.99 for echo intensity) plus close agreement with manual measurements.

What carries the argument

MonoUNet: an aggressively reduced U-Net backbone augmented by a trainable monogenic block that extracts multi-scale local phase features and a gating mechanism that injects them into the encoder stages to reduce sensitivity to ultrasound appearance variations.

If this is right

  • MonoUNet can run in real time on portable POCUS hardware for automated cartilage analysis during routine visits.
  • Automated thickness and echo-intensity outputs match manual values closely enough to support consistent osteoarthritis tracking.
  • The model maintains accuracy across different sites and devices, reducing the need for device-specific retraining.
  • Parameter and compute reductions enable deployment without specialized servers or large memory footprints.

Where Pith is reading between the lines

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

  • The monogenic gating idea could be tested on other compact medical imaging tasks where speckle and intensity variation degrade performance.
  • On-device adaptation of the trainable monogenic block might let clinics fine-tune the model to their specific ultrasound machines without full retraining.
  • Combining the approach with simple mobile apps could enable guided home monitoring of joint cartilage changes between clinic visits.
  • Similar lightweight phase-feature injection may improve segmentation in other ultrasound applications such as tendon or muscle assessment.

Load-bearing premise

The multi-site, multi-device training data plus the specific monogenic block and gating combination are enough to make the model robust to real-world ultrasound variations without overfitting to particular devices or acquisition settings.

What would settle it

Performance measured on an independent set of images from new ultrasound devices or clinics where Dice falls below 85% or Bland-Altman limits of agreement for thickness or echo intensity are exceeded.

Figures

Figures reproduced from arXiv: 2604.07780 by Alvin Kimbowa, Arjun Parmar, David Liu, Ibrahim Mujtaba, Ilker Hacihaliloglu, Matthew Harkey, Maziar Badii, Will Wei.

Figure 1
Figure 1. Figure 1: Overview of MonoUNet architecture: Trainable multi-scale local phase features are extracted from the input image using the Mono block and injected into the high-resolution encoder stages via Mono gates, where they are fused with the encoder features using learnable channel-wise weights. The decoder is half the size of the encoder. via the Mono gate. We detail the different architectural components in the f… view at source ↗
Figure 2
Figure 2. Figure 2: Cross-dataset Dice score distributions for the three training scenarios. [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Bland-Altman plots showing the agreement between cartilage thickness and [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Representative qualitative results on unseen test images. Manual annotations [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗
read the original abstract

Objective: To develop a robust and compact deep learning model for automated knee cartilage segmentation on point-of-care ultrasound (POCUS) devices. Methods: We propose MonoUNet, a novel, highly compact segmentation model consisting of (i) an aggressively reduced U-Net backbone, (ii) a trainable monogenic block that extracts multi-scale local phase features from the input, and (iii) a gating mechanism that injects these features into the encoder stages to reduce sensitivity to variations in ultrasound image appearance. MonoUNet segmentation performance was evaluated on a multi-site, multi-device knee cartilage ultrasound dataset using Dice score and mean average surface distance (MASD). Agreement between MonoUNet and manual cartilage outcomes (thickness and echo intensity) was assessed using Bland-Altman analysis with 95% limits of agreement, and reliability was assessed using intraclass correlation coefficient (ICC$_{2,k}$). Results: Overall, MonoUNet outperformed existing lightweight segmentation models, with average Dice scores ranging from 92.62% to 94.82% and MASD values between 0.133 mm and 0.254 mm. MonoUNet reduces the number of parameters by 10x--700x and computational cost by 14x--2000x relative to existing lightweight models. MonoUNet cartilage outcomes showed excellent reliability and agreement with the manual outcomes: intraclass correlation coefficients (ICC$_{2,k})$=0.96 and bias=2.00% (0.047 mm) for average thickness, and ICC$_{2,k}$=0.99 and bias=0.80% (0.328 a.u.) for echo intensity. Conclusion: Incorporating trainable local phase features improves the robustness of highly compact neural networks for knee cartilage segmentation across varying acquisition settings and could support scalable ultrasound-based assessment and monitoring of knee osteoarthritis using POCUS devices. The code is publicly available at https://github.com/alvinkimbowa/monounet.

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 introduces MonoUNet, a compact segmentation network for knee cartilage on point-of-care ultrasound consisting of an aggressively reduced U-Net backbone, a trainable monogenic block extracting multi-scale local phase features, and a gating mechanism to inject these features into the encoder. On a multi-site, multi-device dataset, it reports Dice scores of 92.62–94.82% and MASD of 0.133–0.254 mm, outperforming other lightweight models while using 10×–700× fewer parameters and 14×–2000× less compute; it also shows high agreement with manual cartilage thickness (ICC 0.96, bias 0.047 mm) and echo intensity (ICC 0.99, bias 0.328 a.u.) via Bland-Altman and ICC analysis.

Significance. If the claimed robustness to acquisition variations holds under proper generalization testing, the work would enable scalable, automated cartilage assessment on portable POCUS devices for knee osteoarthritis monitoring. The public code release at the cited GitHub repository is a clear strength supporting reproducibility and further development.

major comments (2)
  1. [Methods] Methods (dataset and evaluation protocol): The central robustness claim—that the monogenic block plus gating yields invariance to device-specific ultrasound appearance—requires that test images come from acquisition settings (devices, sites, gain/frequency settings) absent from training. The manuscript does not describe whether the multi-site, multi-device split was stratified by device or site (e.g., leave-one-device-out or site-wise hold-out) versus a standard patient-wise random split; without this, the reported Dice/MASD/ICC values cannot be taken as evidence of generalization to new POCUS hardware.
  2. [Results] Results (ablation and statistical support): No ablation study isolating the contribution of the trainable monogenic block versus the gating mechanism or the reduced U-Net alone is presented, nor are dataset size, number of images per device/site, or statistical significance tests (e.g., paired t-tests or Wilcoxon on Dice across models) reported. These omissions make it impossible to verify that the performance gains and “excellent agreement” are attributable to the proposed components rather than dataset characteristics.
minor comments (2)
  1. [Abstract] Abstract: The range “92.62% to 94.82%” for Dice is reported without clarifying whether these are per-fold, per-device, or overall means; adding this clarification would improve interpretability.
  2. [Abstract] Notation: ICC is written as ICC$_{2,k}$ in the abstract but later as ICC(2,k); consistent subscript formatting should be used throughout.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. The comments highlight important aspects of experimental design and reporting that will strengthen the manuscript. We address each major comment below and will revise the paper to incorporate clarifications, additional experiments, and statistical analyses where appropriate.

read point-by-point responses

  1. Referee: [Methods] Methods (dataset and evaluation protocol): The central robustness claim—that the monogenic block plus gating yields invariance to device-specific ultrasound appearance—requires that test images come from acquisition settings (devices, sites, gain/frequency settings) absent from training. The manuscript does not describe whether the multi-site, multi-device split was stratified by device or site (e.g., leave-one-device-out or site-wise hold-out) versus a standard patient-wise random split; without this, the reported Dice/MASD/ICC values cannot be taken as evidence of generalization to new POCUS hardware.

    Authors: We agree that explicit description of the data partitioning is essential to support the robustness claims. The original experiments used a patient-wise random split (80/20 train/test) with no patient overlap, and all devices and sites were represented in both partitions to reflect real-world multi-device variability. This does not constitute a strict leave-one-device-out protocol. In the revised manuscript we will (i) clearly state the patient-wise random split protocol in the Methods section, (ii) report the number of images per device and site, and (iii) add a supplementary leave-one-device-out evaluation (or explicitly discuss its computational feasibility and limitations) to better quantify generalization to unseen hardware. revision: yes

  2. Referee: [Results] Results (ablation and statistical support): No ablation study isolating the contribution of the trainable monogenic block versus the gating mechanism or the reduced U-Net alone is presented, nor are dataset size, number of images per device/site, or statistical significance tests (e.g., paired t-tests or Wilcoxon on Dice across models) reported. These omissions make it impossible to verify that the performance gains and “excellent agreement” are attributable to the proposed components rather than dataset characteristics.

    Authors: We acknowledge these omissions. The revised manuscript will include a dedicated ablation study that isolates the monogenic block, the gating mechanism, and the reduced U-Net backbone. We will also report exact dataset sizes and image counts per device/site, and add paired statistical tests (Wilcoxon signed-rank or paired t-tests with correction) comparing Dice and MASD across models. These additions will allow readers to attribute performance differences to the proposed components. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical evaluation on held-out data is self-contained

full rationale

The paper presents an empirical ML architecture (reduced U-Net + monogenic block + gating) trained and tested on a multi-site ultrasound dataset. Performance claims rest on standard held-out metrics (Dice, MASD, ICC, Bland-Altman) rather than any derivation, first-principles prediction, or parameter fit that reduces to the inputs by construction. No self-definitional equations, fitted-input predictions, or load-bearing self-citations appear in the abstract or described methods. The central result is an experimental outcome, not a tautological renaming or imported uniqueness theorem.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The central claim rests on standard deep learning assumptions plus the effectiveness of the proposed monogenic and gating components, which are validated only through empirical performance on the authors' dataset.

free parameters (1)
  • MonoUNet network weights and hyperparameters
    All model parameters are fitted to the knee cartilage segmentation task on the training data; exact count and specific hyperparameter values are not stated in the abstract.
axioms (1)
  • domain assumption Trainable monogenic features improve robustness to ultrasound appearance variations
    The paper assumes that injecting multi-scale local phase features via the monogenic block and gating will reduce sensitivity to device and acquisition differences.
invented entities (1)
  • Trainable monogenic block no independent evidence
    purpose: To extract multi-scale local phase features from ultrasound input for injection into the encoder
    A new component introduced in the model architecture; no independent evidence outside the empirical results is provided in the abstract.

pith-pipeline@v0.9.0 · 5701 in / 1509 out tokens · 66506 ms · 2026-05-12T01:46:31.343791+00:00 · methodology

Review history (2 revisions) →

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

40 extracted references · 40 canonical work pages · 1 internal anchor

  1. [1]

    A. Cui, H. Li, D. Wang, J. Zhong, Y. Chen, H. Lu, Global, regional prevalence, incidence and risk factors of knee osteoarthri- tis in population-based studies, eClinicalMedicine 29 (2020). URL:https://www.thelancet.com/journals/eclinm/article/ PIIS2589-5370%2820%2930331-X/fulltext?ref=goodoil.news. doi:10.1016/j.eclinm.2020.100587, publisher: Elsevier

    work page doi:10.1016/j.eclinm.2020.100587 2020

  2. [2]

    Y. He, Z. Li, P. G. Alexander, B. D. Ocasio-Nieves, L. Yocum, H. Lin, R. S. Tuan, Pathogenesis of Osteoarthritis: Risk Factors, Regula- tory Pathways in Chondrocytes, and Experimental Models, Biology 9 (2020). URL:https://www.mdpi.com/2079-7737/9/8/194. doi:10. 3390/biology9080194, company: Multidisciplinary Digital Publishing Institute Distributor: Multi...

    work page 2020

  3. [3]

    C. A. Emery, J. L. Whittaker, A. Mahmoudian, L. S. Lohmander, E. M. Roos, K. L. Bennell, C. M. Toomey, R. A. Reimer, D. Thompson, J. L. Ronsky, G. Kuntze, D. G. Lloyd, T. Andriacchi, M. Englund, V. B. Kraus, E. Losina, S. Bierma-Zeinstra, J. Runhaar, G. Peat, F. P. Luyten, L. Snyder-Mackler, M. A. Risberg, A. Mobasheri, A. Guer- mazi, D. J. Hunter, N. K. ...

    work page 2019

  4. [4]

    doi:10.1038/s41584-019-0237-3, publisher: Nature Publishing Group

    URL:https://www.nature.com/articles/s41584-019-0237-3. doi:10.1038/s41584-019-0237-3, publisher: Nature Publishing Group. 18

    work page doi:10.1038/s41584-019-0237-3

  5. [5]

    J. E. Collins, P. Mesenbrink, R. Jin, E. B. Dam, L. A. Deveza, F. Eck- stein, A. Guermazi, C. Ladel, T. A. Perry, D. Robinson, F. W. Roemer, C. J. Swearingen, W. Wirth, V. B. Kraus, D. J. Hunter, Magnetic Res- onance Imaging Biomarkers of Knee Osteoarthritis Progression, ACR Open Rheumatology 7 (2025) e70085. URL:https://pmc.ncbi.nlm. nih.gov/articles/PMC...

    work page doi:10.1002/acr2.70085 2025

  6. [6]

    E. H. G. Oei, J. Hirvasniemi, T. A. v. Zadelhoff, R. A. v. d. Heijden, Osteoarthritis year in review 2021: imaging, Os- teoarthritis and Cartilage 30 (2022) 226–236. URL:https://www. oarsijournal.com/article/S1063-4584(21)00974-2/fulltext. doi:10.1016/j.joca.2021.11.012, publisher: Elsevier

    work page doi:10.1016/j.joca.2021.11.012 2021

  7. [7]

    F. W. Roemer, A. Guermazi, S. Demehri, W. Wirth, R. Kijowski, Imaging in Osteoarthritis, Osteoarthritis and Cartilage 30 (2022) 913–

    work page 2022

  8. [8]

    doi:10.1016/j.joca.2021.04.018

    URL:https://www.sciencedirect.com/science/article/pii/ S1063458421008815. doi:10.1016/j.joca.2021.04.018

    work page doi:10.1016/j.joca.2021.04.018 2021

  9. [9]

    Nakashima, T

    Y. Nakashima, T. Sunagawa, R. Shinomiya, A. Kodama, N. Adachi, Point-of-care ultrasound in musculoskeletal field, Journal of Medi- cal Ultrasonics 49 (2022) 663–673. URL:https://doi.org/10.1007/ s10396-022-01252-0. doi:10.1007/s10396-022-01252-0

    work page doi:10.1007/s10396-022-01252-0 2022

  10. [10]

    M. S. Harkey, N. Michel, C. Grozier, J. M. Slade, K. Collins, B. Pietrosimone, D. Lalush, C. Lisee, I. Hacihaliloglu, R. Fa- jardo, Femoral cartilage ultrasound echo-intensity is a valid measure of cartilage composition, Journal of Orthopaedic Re- search 42 (2024) 729–736. URL:https://onlinelibrary.wiley. com/doi/abs/10.1002/jor.25722. doi:10.1002/jor.257...

    work page doi:10.1002/jor.25722 2024

  11. [11]

    D’Agostino, A

    V. D’Agostino, A. Sorriento, A. Cafarelli, D. Donati, N. Papalexis, A. Russo, G. Lisignoli, L. Ricotti, P. Spinnato, Ultrasound Imaging in Knee Osteoarthritis: Current Role, Recent Advancements, and Future Perspectives, Journal of Clinical Medicine 13 (2024) 4930. URL:https: //www.mdpi.com/2077-0383/13/16/4930. doi:10.3390/jcm13164930, number: 16 Publishe...

    work page doi:10.3390/jcm13164930 2024

  12. [12]

    A. S. Parmar, C. D. Grozier, J. E. Tolzman, R. Dima, B. Winn, I. Hacihaliloglu, R. Fajardo, M. S. Harkey, Wireless Ultrasound 19 Probes: A New Frontier In Assessing Femoral Cartilage Health: 249, Medicine & Science in Sports & Exercise 56 (2024) 74. URL: https://journals.lww.com/acsm-msse/fulltext/2024/10001/ wireless_ultrasound_probes__a_new_frontier_in....

    work page doi:10.1249/01.mss.0001053124.19355.22 2024

  13. [13]

    Desai, I

    P. Desai, I. Hacihaliloglu, Knee-Cartilage Segmentation and Thick- ness Measurement from 2D Ultrasound, Journal of Imaging 5 (2019)

    work page 2019

  14. [14]

    doi:10.3390/ jimaging5040043, number: 4 Publisher: Multidisciplinary Digital Pub- lishing Institute

    URL:https://www.mdpi.com/2313-433X/5/4/43. doi:10.3390/ jimaging5040043, number: 4 Publisher: Multidisciplinary Digital Pub- lishing Institute

    work page

  15. [15]

    Zhang, E

    H. Zhang, E. Ning, L. Lu, J. Zhou, Z. Shao, X. Yang, Y. Hao, Research progress of ultrasound in accurate evaluation of car- tilage injury in osteoarthritis, Frontiers in Endocrinology 15 (2024). URL:https://www.frontiersin.org/journals/ endocrinology/articles/10.3389/fendo.2024.1420049/full. doi:10.3389/fendo.2024.1420049, publisher: Frontiers

    work page doi:10.3389/fendo.2024.1420049/full 2024

  16. [16]

    M. S. Harkey, N. Michel, C. Kuenze, R. Fajardo, M. Salzler, J. B. Driban, I. Hacihaliloglu, Validating a semi-automated technique for segmenting femoral articular cartilage on ultrasound images, Cartilage 13 (2022) 19476035221093069

    work page 2022

  17. [17]

    Papernick, R

    S. Papernick, R. Dima, D. J. Gillies, C. T. Appleton, A. Fen- ster, Reliability and concurrent validity of three-dimensional ultra- sound for quantifying knee cartilage volume, Osteoarthritis and Car- tilage Open 2 (2020). URL:https://www.oarsiopenjournal.com/ article/S2665-9131(20)30127-8/fulltext. doi:10.1016/j.ocarto. 2020.100127, publisher: Elsevier

    work page doi:10.1016/j.ocarto 2020

  18. [18]

    Salimi, A

    N. Salimi, A. Gonzalez-Fiol, D. Yanez, K. Fardelmann, E. Har- mon, K. Kohari, S. Abdel-Razeq, U. Magriples, A. Alian, Ultra- sound Image Quality Comparison Between a Handheld Ultrasound Transducer and Mid-Range Ultrasound Machine: Image character- istics of the Butterfly iQ vs. Sonosite M-Turbo, POCUS Journal 7 (2022) 154–159. URL:https://ojs.library.quee...

    work page doi:10.24908/pocus.v7i1.15052 2022

  19. [19]

    Perez-Sanchez, G

    A. Perez-Sanchez, G. Johnson, N. Pucks, R. N. Soni, T. J. S. Lund, A. J. Andrade, M.-P. T. Le, J. Solis-McCarthy, T. Wong, A. Ashraf, A. D. Kumar, G. I. Banauch, J. R. Verner, A. Sodhi, M. K. Thomas, C. Lo- Presti, H. Schmitz, A. Koratala, J. Hunninghake, E. Manninen, C. Can- dotti, T. Minami, B. K. Mathews, G. Bandak, H. Sauthoff, H. Mayo- Malasky, J. Ch...

    work page doi:10.1186/s13089-024-00392-3 2024

  20. [20]

    URL:https://pmc.ncbi.nlm.nih.gov/articles/ PMC6746898/

    ESR statement on portable ultrasound devices, Insights into Imag- ing 10 (2019) 89. URL:https://pmc.ncbi.nlm.nih.gov/articles/ PMC6746898/. doi:10.1186/s13244-019-0775-x

    work page doi:10.1186/s13244-019-0775-x 2019

  21. [21]

    Rykkje, J

    A. Rykkje, J. F. Carlsen, M. B. Nielsen, Hand-Held Ultrasound De- vices Compared with High-End Ultrasound Systems: A Systematic Re- view, Diagnostics 9 (2019) 61. URL:https://pmc.ncbi.nlm.nih.gov/ articles/PMC6628329/. doi:10.3390/diagnostics9020061

    work page doi:10.3390/diagnostics9020061 2019

  22. [22]

    M. M. Ahmed, O. J. Okesanya, N. O. Olaleke, O. A. Adigun, U. O. Adebayo, T. A. Oso, G. Eshun, D. E. Lucero-Prisno, Integrating Digi- tal Health Innovations to Achieve Universal Health Coverage: Promot- ing Health Outcomes and Quality Through Global Public Health Eq- uity, Healthcare 13 (2025) 1060. URL:https://pmc.ncbi.nlm.nih. gov/articles/PMC12071628/. ...

    work page doi:10.3390/healthcare13091060 2025

  23. [23]

    A. D. Desai, F. Caliva, C. Iriondo, A. Mortazi, S. Jambawalikar, U. Bagci, M. Perslev, C. Igel, E. B. Dam, S. Gaj, M. Yang, X. Li, C. M. Deniz, V. Juras, R. Regatte, G. E. Gold, B. A. Hargreaves, V. Pedoia, A. S. Chaudhari, N. Khosravan, D. Torigian, J. Ellermann, M. Ak- cakaya, R. Tibrewala, I. Flament, M. O’Brien, S. Majumdar, K. Naka- mura, A. Pai, The...

    work page doi:10.1148/ryai.2021200078 2021

  24. [24]

    Antico, F

    M. Antico, F. Sasazawa, M. Dunnhofer, S. M. Camps, A. T. Jaiprakash, A. K. Pandey, R. Crawford, G. Carneiro, D. Fonta- narosa, Deep Learning-Based Femoral Cartilage Automatic Seg- mentation in Ultrasound Imaging for Guidance in Robotic Knee Arthroscopy, Ultrasound in Medicine and Biology 46 (2020) 422–

    work page 2020

  25. [25]

    doi:10.1016/j.ultrasmedbio.2019.10.015, pub- lisher: Elsevier

    URL:https://www.umbjournal.org/article/S0301-5629(19) 31552-2/fulltext. doi:10.1016/j.ultrasmedbio.2019.10.015, pub- lisher: Elsevier

    work page doi:10.1016/j.ultrasmedbio.2019.10.015 2019

  26. [26]

    J. M. J. Valanarasu, V. M. Patel, UNeXt: MLP-Based Rapid Medical Image Segmentation Network, in: L. Wang, Q. Dou, P. T. Fletcher, S. Speidel, S. Li (Eds.), Medical Image Computing and Computer Assisted Intervention – MICCAI 2022, Springer Nature Switzerland, Cham, 2022, pp. 23–33. doi:10.1007/978-3-031-16443-9_3

    work page doi:10.1007/978-3-031-16443-9_3 2022

  27. [27]

    Kalkhof, C

    J. Kalkhof, C. González, A. Mukhopadhyay, Med-NCA: Ro- bust and Lightweight Segmentation with Neural Cellular Au- tomata, 2023. URL:http://arxiv.org/abs/2302.03473. doi:10. 48550/arXiv.2302.03473, arXiv:2302.03473 [cs, eess]

    work page arXiv 2023

  28. [28]

    J. Chen, R. Chen, W. Wang, J. Cheng, L. Zhang, L. Chen, TinyU- Net: Lighter Yet Better U-Net with Cascaded Multi-receptive Fields, in: M. G. Linguraru, Q. Dou, A. Feragen, S. Giannarou, B. Glocker, K. Lekadir, J. A. Schnabel (Eds.), Medical Image Computing and Com- puter Assisted Intervention – MICCAI 2024, Springer Nature Switzer- land, Cham, 2024, pp. 6...

    work page doi:10.1007/978-3-031-72114-4_60 2024

  29. [29]

    Hacihaliloglu, R

    I. Hacihaliloglu, R. Abugharbieh, A. J. Hodgson, R. N. Rohling, 2A- 4 Enhancement of Bone Surface Visualization from 3D Ultrasound Based on Local Phase Information, in: 2006 IEEE Ultrasonics Symposium, 2006, pp. 21–24. URL:https://ieeexplore.ieee.org/ document/4151876. doi:10.1109/ULTSYM.2006.20, iSSN: 1051-0117

    work page doi:10.1109/ultsym.2006.20 2006

  30. [30]

    Hacihaliloglu, R

    I. Hacihaliloglu, R. Abugharbieh, A. J. Hodgson, R. N. Rohling, Bone Surface Localization in Ultrasound Using Image Phase-Based Features, Ultrasound in Medicine & Biology 35 (2009) 1475–1487. URL:https://www.sciencedirect.com/science/article/pii/ S030156290900177X. doi:10.1016/j.ultrasmedbio.2009.04.015. 22

    work page doi:10.1016/j.ultrasmedbio.2009.04.015 2009

  31. [31]

    nnU-Net: A Self-Configuring Method for Deep Learning-Based Biomedical Image Segmentation,

    F. Isensee, P. F. Jaeger, S. A. A. Kohl, J. Petersen, K. H. Maier-Hein, nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation, Nature Meth- ods 18 (2021) 203–211. URL:https://www.nature.com/articles/ s41592-020-01008-z. doi:10.1038/s41592-020-01008-z, publisher: Nature Publishing Group

    work page doi:10.1038/s41592-020-01008-z 2021

  32. [32]

    arXiv preprint arXiv:2404.09556 , year=

    F. Isensee, T. Wald, C. Ulrich, M. Baumgartner, S. Roy, K. Maier-Hein, P. F. Jaeger, nnU-Net Revisited: A Call for Rigorous Validation in 3D Medical Image Segmentation, 2024. URL:http://arxiv.org/abs/ 2404.09556. doi:10.48550/arXiv.2404.09556, arXiv:2404.09556 [cs]

    work page doi:10.48550/arxiv.2404.09556 2024

  33. [33]

    https://doi.org/10.48550/arXiv.2512.03834, https://arxiv.org/abs/2512.03834

    T. Hassler, I. Åkerholm, M. Nordström, G. Balletti, O. Gok- sel, Lean Unet: A Compact Model for Image Segmenta- tion, 2025. URL:http://arxiv.org/abs/2512.03834. doi:10.48550/ arXiv.2512.03834, arXiv:2512.03834 [cs]

    work page arXiv 2025

  34. [34]

    E. U. Moya-Sánchez, S. Xambó-Descamps, A. S. Pérez, S. Salazar- Colores, U. Cortés, A Trainable Monogenic ConvNet Layer Robust in Front of Large Contrast Changes in Image Classification, IEEE Access 9 (2021) 163735–163746. URL:https://ieeexplore.ieee. org/document/9615499. doi:10.1109/ACCESS.2021.3128552, confer- ence Name: IEEE Access

    work page doi:10.1109/access.2021.3128552 2021

  35. [35]

    Hacihaliloglu, R

    I. Hacihaliloglu, R. Abugharbieh, A. J. Hodgson, R. N. Rohling, Automatic Adaptive Parameterization in Local Phase Feature- Based Bone Segmentation in Ultrasound, Ultrasound in Medicine & Biology 37 (2011) 1689–1703. URL:https://www. sciencedirect.com/science/article/pii/S0301562911011744. doi:10.1016/j.ultrasmedbio.2011.06.006

    work page doi:10.1016/j.ultrasmedbio.2011.06.006 2011

  36. [36]

    Decoupled Weight Decay Regularization

    I. Loshchilov, F. Hutter, Decoupled Weight Decay Regulariza- tion, 2019. URL:http://arxiv.org/abs/1711.05101. doi:10.48550/ arXiv.1711.05101, arXiv:1711.05101 [cs]

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  37. [37]

    R. Azad, M. Heidary, K. Yilmaz, M. Hüttemann, S. Karimijafarbigloo, Y. Wu, A. Schmeink, D. Merhof, Loss Functions in the Era of Semantic Segmentation: A Survey and Outlook, 2023. URL:http://arxiv.org/ abs/2312.05391. doi:10.48550/arXiv.2312.05391, arXiv:2312.05391 [cs] version: 1. 23

    work page doi:10.48550/arxiv.2312.05391 2023

  38. [38]

    URL:https://arxiv.org/abs/2308.01239v2

    F.Tang, J.Ding, L.Wang, C.Ning, S.K.Zhou, CMUNeXt: AnEfficient Medical Image Segmentation Network based on Large Kernel and Skip Fusion, 2023. URL:https://arxiv.org/abs/2308.01239v2

    work page arXiv 2023

  39. [39]

    Paszke, S

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, S. Chintala, PyTorch: An Imperative Style, High-Performance Deep Learning Library, in: Advances in Neu- ral Information Processing Syst...

    work page 2019

  40. [40]

    Maier-Hein, A

    L. Maier-Hein, A. Reinke, P. Godau, M. D. Tizabi, F. Buettner, E. Christodoulou, B. Glocker, F. Isensee, J. Kleesiek, M. Kozubek, M. Reyes, M. A. Riegler, M. Wiesenfarth, A. E. Kavur, C. H. Su- dre, M. Baumgartner, M. Eisenmann, D. Heckmann-Nötzel, T. Räd- sch, L. Acion, M. Antonelli, T. Arbel, S. Bakas, A. Benis, M. B. Blaschko, M. J. Cardoso, V. Cheplyg...

    work page doi:10.1038/s41592-023-02151-z 2024