Pith. sign in

REVIEW 2 major objections 5 minor 31 references

A multi-decoder segmentation network pre-trained on cholecystectomy videos transfers to rectal-surgery images, reaching 62.4% dice after full fine-tuning and matching a from-scratch model (60.0%) with no fine-tuning.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 05:35 UTC pith:3LBKO52U

load-bearing objection Useful, incremental empirical work on decoder-specific transfer for laparoscopic segmentation, but the zero-shot transfer claim is under-specified to the point of being unverifiable as written. the 2 major comments →

arxiv 2607.29509 v1 pith:3LBKO52U submitted 2026-07-31 cs.CV cs.LG

Leveraging Transfer Learning with Class-Specific Decoders for Laparoscopic Segmentation

Priya Tomar , Aditya Parikh , Christian Bauckhage , Rafet Sifa This is my paper
classification cs.CV cs.LG
keywords laparoscopic segmentationmulti-organ segmentationclass-specific decoderstransfer learningCEMDCholecSeg8KDresden Surgical Anatomyclass imbalance
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper sets out to show that supervised transfer learning works between two different laparoscopic procedures — from cholecystectomy (gallbladder surgery) to rectal surgery — when the segmentation model keeps a separate decoder head per organ. On the target procedure's dataset, the best configuration (CEMD FFT) reaches 62.4% dice, beating the 60.0% from-scratch baseline while converging roughly three times faster. Even with zero fine-tuning, the transferred model ties the from-scratch score at about 60% dice, which the paper reads as evidence of genuine cross-surgical anatomical knowledge transfer. The paper also isolates where that knowledge is stored: freezing the encoder and tuning only decoders consistently hurts, whereas full-network fine-tuning helps, so the encoder must adapt to the new domain. If these results hold, annotated surgical video from one procedure type could bootstrap segmentation in another, easing the annotation bottleneck.

Core claim

The paper claims that a Common Encoder–Multiple Decoder (CEMD) architecture, pre-trained on CholecSeg8K and fully fine-tuned on the Dresden Surgical Anatomy (DSA) dataset, achieves the best segmentation performance at 62.4% dice, with faster convergence than training from scratch. It further claims that the same pre-trained model evaluated on DSA with no fine-tuning reaches about 60% dice, comparable to a model trained on DSA from scratch, demonstrating effective cross-surgical knowledge transfer despite only partial class overlap. On the source dataset, CEMD outperforms a shared-decoder baseline by about 4% overall dice, with a 27-point gain on the underrepresented Gallbladder class. Fine-t

What carries the argument

The central object is the Common Encoder–Multiple Decoder (CEMD) architecture: a shared encoder (an Attention U-Net, i.e., a U-Net with attention gates) with a dedicated decoder head for each organ class, contrasted with a Common Encoder–Common Decoder (CECD) baseline. The class-specific decoders are the mechanism that lets the model separate anatomical features per organ; the transfer protocol (CholecSeg8K pre-training followed by DSA fine-tuning with the encoder either frozen or trainable) is what exposes where transferable knowledge lives.

Load-bearing premise

The load-bearing premise is that the eleven DSA decoder heads used in the no-fine-tuning evaluation have meaningful initializations from CholecSeg8K pre-training; the paper reports dice for DSA-only organs without stating how those decoder heads are initialized or which CholecSeg8K class maps to them.

What would settle it

Reset the decoder heads for DSA-only organs (Spleen, Ureter, etc.) to random weights and run the same zero-shot evaluation on DSA; if the overall dice remain near 60%, the transfer result is not coming from learned organ-specific decoders. Inspecting the pre-trained checkpoint to confirm each of the eleven decoder heads was actually updated on CholecSeg8K would settle the same question.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Pre-training on one laparoscopic procedure type can substitute for training from scratch on another, at equal or better accuracy and with substantially faster convergence.
  • Full-network fine-tuning is necessary to realize transfer gains; freezing the encoder and updating only decoders is worse than doing no fine-tuning at all.
  • Organ-specific decoder heads transfer better than a shared decoder under domain shift, losing 4.9 versus 5.6 dice points when the encoder is frozen.
  • Class-specific decoders concentrate their benefit on rare classes (Gallbladder +27% dice on CholecSeg8K), but transfer learning does not fix class imbalance for the smallest organs on DSA.
  • The zero-shot transfer result, if reliable, means target-domain labels are not strictly required to obtain a usable segmentation model for a new surgical domain.

Where Pith is reading between the lines

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

  • The zero-shot 60% dice figure is only interpretable if all eleven DSA decoder heads have defined initializations from CholecSeg8K; the paper's 6-class source evaluation leaves open how decoders for DSA-only organs are initialized, so a repeatable experiment should spell that mapping out.
  • A natural extension the authors do not test: pre-train on multiple source procedures jointly and measure whether the zero-shot floor rises above 60% dice before any fine-tuning.
  • The DFT degradation suggests a testable mechanism — shared decoders entangle class representations with the encoder's feature distribution, so a frozen encoder plus retrained decoder misaligns; a feature-distribution shift analysis before and after DFT could confirm this.
  • For deployment, the 0FT result hints at a label-efficient workflow: generate initial segmentations with the zero-shot model, have surgeons correct only the low-confidence regions, then fine-tune; this is an editorial extrapolation, not reported in the paper.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper studies transfer learning between two laparoscopic datasets (CholecSeg8K and DSA) for multi-organ segmentation under two architectures: a shared-decoder baseline (CECD) and a class-specific multi-decoder architecture (CEMD). It reports that CEMD outperforms CECD on both datasets, that a CholecSeg8K-pretrained CEMD evaluated on DSA without fine-tuning reaches roughly 60% Dice (comparable to training from scratch), that full fine-tuning improves this to 62.4% Dice, and that decoder-only fine-tuning degrades performance relative to no fine-tuning. The paper interprets these results as evidence of effective cross-surgical knowledge transfer and argues that class-specific decoders retain transferred knowledge better. It also emphasizes that class imbalance remains a persistent bottleneck.

Significance. If the experimental protocol is made precise, the paper would provide a useful empirical data point: supervised transfer between two different laparoscopic procedure domains under a class-specific decoder architecture, with a clean separation of encoder-only versus decoder-only adaptation. The paper is transparent about building on the authors' earlier CEMD design [17] and re-baselines it, so circularity concerns are low. It reports held-out test Dice/IoU, includes convergence curves, and candidly discusses limitations. The main risk is not circularity but incomplete specification of the zero-shot evaluation protocol: the CholecSeg8K model is trained on a 6-class subset, yet Tables 2-4 report zero-shot Dice for all 11 DSA classes without stating a decoder mapping or initialization. Because the zero-shot result is the paper's second headline contribution, this gap is substantive.

major comments (2)
  1. [§4.1.2, Tables 2-4] The central zero-shot transfer claim is not reproducible as written. CholecSeg8K pre-training is described for a 6-class subset (Background, Abdominal Wall, Liver, Fat, Gallbladder, L-hook), but Tables 2-4 report 0FT Dice/IoU for all 11 DSA classes. Section 3.1 notes that the Cholec gastrointestinal class combines stomach and small intestine, but it does not specify how Colon, Pancreas, Spleen, Ureter, Vesicular Glands, Inferior Mesenteric Artery, or Intestinal Veins are represented by the pre-trained CEMD. The paper must state: (i) how many decoders the CholecSeg8K CEMD has; (ii) how the decoders for DSA classes absent from CholecSeg8K are initialized for 0FT (random, copied, or mapped from a related class); and (iii) the exact class mapping used to produce the per-class 0FT numbers. Without this, the conclusion that CEMD 0FT is comparable to training from scratch (Table 2: 60.0 vs. 60.
  2. [§3.2, §4.2.1, Tables 3-4] The decoder fine-tuning (DFT) comparison has the same ambiguity and affects the encoder/decoder adaptation claim. If CEMD is instantiated with 11 decoders during CholecSeg8K training, it is unclear which decoders receive gradients from the 6 training classes. If it is instantiated with 6 decoders, it is unclear how the 11 DSA decoders are created for DFT/FFT. The claim that CEMD is more resilient to DFT degradation than CECD (−4.9% vs. −5.6%) could conflate freshly initialized decoders for unseen DSA organs with transferred decoders. The paper should specify the number of decoders in each phase and report which decoders were pre-trained, so the resilience comparison is interpretable.
minor comments (5)
  1. [§4.1.1, Table 5, Conclusion] The text says CEMD improves overall Dice by about 3.5% over CECD when trained from scratch on DSA, but Table 5 reports 60.0 vs. 59.3, a 0.7-point gain. If the 3.5% refers to per-class average gain or to a different baseline from [17], state that explicitly. As written, the numbers conflict.
  2. [Table 2] The Best column is not the maximum of the three reported epoch columns: CEMD 0FT shows 60.7 at epoch 150 but Best = 60.0. If Best means the checkpoint with lowest validation loss rather than highest test Dice, say so; the current presentation is confusing.
  3. [Tables 3-4] The notation uses a leading plus/minus both for the DFT/FFT values and for the change relative to 0FT (e.g., −85.52−75.13). This makes the tables hard to parse. Please reformat, e.g., give 0FT, DFT, FFT as separate columns with delta columns clearly labeled.
  4. [§3.3, §4.2.1] The phrase No Fine-Tuning (0FT) is used both for training from scratch on the same dataset and for cross-dataset zero-shot evaluation. This dual meaning makes sentences like the scores are in a similar range to the in-dataset evaluation on the DSA dataset ambiguous. Use distinct labels, e.g., scratch vs. zero-shot transfer.
  5. [§3.3, Limitations] All results are from a single seed with no confidence intervals or significance tests. Given that several headline differences are small (e.g., CECD FFT vs. CECD 0FT: 59.4 vs. 59.3), the paper should at least acknowledge the lack of variance quantification.

Circularity Check

0 steps flagged

No significant circularity: the paper's claims are measured benchmark results, and the self-citation to the authors' prior CEMD work is corroborative rather than load-bearing.

full rationale

The central claims are empirical: the 0FT, DFT, and FFT dice scores in Tables 2-4 are measured on a held-out DSA test set, not derived from parameters fitted to DSA and renamed as predictions. No equation in the paper identifies a derived quantity with an input. The CEMD-vs-CECD advantage is independently re-established in this paper on DSA (§4.1.1) and CholecSeg8K (§4.1.2, Table 1), so the authors' prior work [17] is used for architectural motivation and consistency, not as an unverified premise that forces the new results. The 'without any fine-tuning... comparable to training from scratch' claim is a direct evaluation of a CholecSeg8K-trained model on DSA, so it is not a fit-to-target wrapped as a prediction. A genuine reproducibility gap exists: §4.1.2 says CholecSeg8K pre-training used six selected classes, while §3.2 and Tables 3-4 report 0FT dice for eleven DSA classes without specifying how the additional CEMD decoders are initialized or mapped from Cholec classes. That is a validity/reporting concern for the 0FT transfer claim, but it does not make the result equivalent to its own inputs by construction. There is no uniqueness theorem, no ansatz smuggled via citation, and no renaming of a known result; therefore the circularity score is 0.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The ledger shows the hand-set hyperparameters and the under-specified class-alignment assumption on which the reported numbers rest.

free parameters (6)
  • initial learning rate = 1e-4
    Set by hand (§3.3), no sweep reported; convergence and final dice depend on it.
  • learning-rate decay schedule = 0.5 every 10 epochs
    Hand-chosen schedule (§3.3); influences the reported epoch-level convergence claims.
  • max epochs / early stopping = 150 epochs, patience 10
    Hand-chosen (§3.3); the 'faster convergence' comparison is defined relative to this schedule.
  • DSA test-split class proportions = kept high and close to training proportions
    Deliberately chosen in §3.3 and Figures 2-3; absolute dice are conditional on this test distribution.
  • CholecSeg8K class subset = 6 of 13 classes (Background, Abdominal Wall, Liver, Fat, Gallbladder, L-hook)
    Selected in §4.1.2 for 'reliable insights'; excludes sparse classes and shapes the Cholec baseline scores.
  • random seed = fixed but not disclosed
    All runs use 'a fixed seed' (§3.3); with one seed, dice differences have no error estimate.
axioms (4)
  • domain assumption DSA and CholecSeg8K ground-truth annotations are reliable and their class labels can be related for transfer (Liver, Abdominal Wall, and gastrointestinal structures).
    Adopted from dataset papers [12,13] and used in §3.1/§4.2.1; if labels or correspondences are wrong, the transfer conclusion changes.
  • ad hoc to paper A model trained on a 6-class CholecSeg8K subset can be evaluated zero-shot on 11 DSA classes without an explicit decoder alignment.
    Needed for CEMD 0FT in Tables 2-4, but the mechanism is not described anywhere in the paper (see §4.1.2 and §4.2.1).
  • domain assumption Attention U-Net with dice loss and Adam, trained with the stated hyperparameters, is representative of surgical segmentation models.
    The paper uses a single backbone (§3.2) and acknowledges in §5 that results may not generalize to transformers or other architectures.
  • domain assumption The selected test splits are representative of each dataset's class distribution aside from class imbalance.
    The split choice is described as 'we tried to keep the proportion of the test data high and within the range of the training data' (§3.3), suggesting curation.

pith-pipeline@v1.3.0-daily-deepseek · 11140 in / 21382 out tokens · 203598 ms · 2026-08-03T05:35:15.113544+00:00 · methodology

0 comments
read the original abstract

Effective multi-organ segmentation in surgical data requires learning the intricate anatomical features and alleviating the challenge of class imbalance, which results from relatively lower proportions of small and limitedly exposed structures. Recent works on laparoscopic multi-organ segmentation focus on learning structure-specific features through class-specific decoder architectures and report favorable results. This work extends the decoder-focused architectures to investigate knowledge sharing in the cross-surgical domain. We utilize two datasets representing different surgical domains, rectal and cholecystectomy surgeries, to explore how surgical conceptual knowledge transfers under partially common anatomical representations. Additionally, we compare the feature adaptation for the encoder and decoder at different training stages to analyse the knowledge adaptation and retention in the network. Our results corroborate previous findings on decoder-specific architectures and demonstrate that the organ-specific decoder model (CEMD), fully fine-tuned after cross-domain pre-training, achieves the highest segmentation performance (62.4\% dice) while converging substantially faster than training from scratch. However, we also find that class imbalance in surgical data remains a persistent challenge that transfer learning does not fully resolve for underrepresented anatomical structures.

Figures

Figures reproduced from arXiv: 2607.29509 by Aditya Parikh, Christian Bauckhage, Priya Tomar, Rafet Sifa.

Figure 1
Figure 1. Figure 1: Comparison of (top) Common Encoder-Common Decoder (CECD): A single pipeline [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Pixel distribution of Dresden Surgical Anatomy (DSA) dataset classes in training and [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Pixel distribution of CholecSeg8K dataset classes in training and evaluation splits. The [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The difference in dice scores between the Common Encoder Multiple Decoder (CEMD) [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of DICE scores (%) on the validation set across training epochs for the CholecSeg8K dataset using two architectural ap￾proaches, CECD and CEMD. Besides resulting in better performance, the class-specific decoder framework converges faster than the common de￾coder framework CECD. We compare the learning curve of the two frameworks, CECD and CEMD, on a valida￾tion set at different training stages … view at source ↗
Figure 6
Figure 6. Figure 6: Difference in the dice scores for the Common Encoder Common Decoder (CECD) and [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of DICE scores (%) on the validation set across training epochs for training [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗

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

31 extracted references · 1 canonical work pages

  1. [1]

    Vision techniques for anatomical structures in laparoscopic surgery: a comprehen- sive review.Frontiers in Surgery, 12:1557153, 2025

    Ru Zhou, Dan Wang, Hanwei Zhang, Ying Zhu, Lijun Zhang, Tianxiang Chen, Wenqiang Liao, and Zi Ye. Vision techniques for anatomical structures in laparoscopic surgery: a comprehen- sive review.Frontiers in Surgery, 12:1557153, 2025

    2025

  2. [2]

    Tobias Rueckert, Daniel Rueckert, and Christoph Palm. Methods and datasets for segmenta- tion of minimally invasive surgical instruments in endoscopic images and videos: A review of the state of the art.Computers in Biology and Medicine, 169:107929, 2024

    2024

  3. [3]

    Deep learning for surgical instrument recognition and segmentation in robotic- assisted surgeries: a systematic review.Artificial Intelligence Review, 58(1):1, 2024

    Fatimaelzahraa Ali Ahmed, Mahmoud Yousef, Mariam Ali Ahmed, Hasan Omar Ali, Anns Mahboob, Hazrat Ali, Zubair Shah, Omar Aboumarzouk, Abdulla Al Ansari, and Shidin 15 Balakrishnan. Deep learning for surgical instrument recognition and segmentation in robotic- assisted surgeries: a systematic review.Artificial Intelligence Review, 58(1):1, 2024

    2024

  4. [4]

    Augmenting efficient real-time surgical instrument segmentation in video with point tracking and segment anything

    Zijian Wu, Adam Schmidt, Peter Kazanzides, and Septimiu E Salcudean. Augmenting efficient real-time surgical instrument segmentation in video with point tracking and segment anything. Healthcare Technology Letters, 12(1):e12111, 2025

    2025

  5. [5]

    Segmatch: semi-supervised surgical instrument segmentation.Scientific Reports, 15(1):14042, 2025

    Meng Wei, Charlie Budd, Luis C Garcia-Peraza-Herrera, Reuben Dorent, Miaojing Shi, and Tom Vercauteren. Segmatch: semi-supervised surgical instrument segmentation.Scientific Reports, 15(1):14042, 2025

    2025

  6. [6]

    D. K. Venkatesh, D. Rivoir, M. Pfeiffer, F. Kolbinger, and S. Speidel. Data augmentation for surgical scene segmentation with anatomy-aware diffusion models. In2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2280–2290, 2025. doi: 10.1109/W ACV61041.2025.00228

    arXiv 2025

  7. [7]

    Towards more precise automatic analysis: a comprehensive survey of deep learning-based multi-organ segmentation, 2023

    Xiaoyu Liu, Linhao Qu, Ziyue Xie, Jiayue Zhao, Yonghong Shi, and Zhijian Song. Towards more precise automatic analysis: a comprehensive survey of deep learning-based multi-organ segmentation, 2023

    2023

  8. [8]

    Mosmos: Multi-organ segmentation facilitated by medical report supervision.Biomedical Signal Processing and Control, 106:107743, 2025

    Weiwei Tian, Xinyu Huang, Junlin Hou, Caiyue Ren, Longquan Jiang, Rui-Wei Zhao, Gang Jin, Yuejie Zhang, and Daoying Geng. Mosmos: Multi-organ segmentation facilitated by medical report supervision.Biomedical Signal Processing and Control, 106:107743, 2025. ISSN 1746-8094

    2025

  9. [9]

    M¨ uller-Stich, Martin Wagner, and Franziska Mathis-Ullrich

    Paul Maria Scheikl, Stefan Laschewski, Anna Kisilenko, Tornike Davitashvili, Benjamin M¨ uller, Manuela Capek, Beat P. M¨ uller-Stich, Martin Wagner, and Franziska Mathis-Ullrich. Deep learning for semantic segmentation of organs and tissues in laparoscopic surgery.Current Directions in Biomedical Engineering, 6(1):20200016, 2020

    2020

  10. [10]

    Evolution of multiorgan segmentation techniques from traditional to deep learning in abdominal ct images – a systematic review

    Harinder Kaur, Navjot Anttal, and Nirvair Neeru. Evolution of multiorgan segmentation techniques from traditional to deep learning in abdominal ct images – a systematic review. Displays, 73:102223, 04 2022. doi: 10.1016/j.displa.2022.102223

    arXiv 2022

  11. [11]

    A review of deep learning based methods for medical image multi-organ segmentation.Physica Medica, 85:107–122, 2021

    Yabo Fu, Yang Lei, Tonghe Wang, Walter J Curran, Tian Liu, and Xiaofeng Yang. A review of deep learning based methods for medical image multi-organ segmentation.Physica Medica, 85:107–122, 2021

    2021

  12. [12]

    Cholecseg8k: a semantic segmentation dataset for laparoscopic cholecystectomy based on cholec80.arXiv preprint arXiv:2012.12453, 2020

    W-Y Hong, C-L Kao, Y-H Kuo, J-R Wang, W-L Chang, and C-S Shih. Cholecseg8k: a semantic segmentation dataset for laparoscopic cholecystectomy based on cholec80.arXiv preprint arXiv:2012.12453, 2020

    Pith/arXiv arXiv 2012

  13. [13]

    The dresden surgical anatomy dataset for abdominal organ segmentation in surgical data science.Scientific Data, 10, 01 2023

    Matthias Carstens, Franziska Rinner, Sebastian Bodenstedt, Alexander Jenke, J¨ urgen Weitz, et al. The dresden surgical anatomy dataset for abdominal organ segmentation in surgical data science.Scientific Data, 10, 01 2023. doi: 10.1038/s41597-022-01719-2. 16

    doi:10.1038/s41597-022-01719-2 2023

  14. [14]

    Strategies to improve real-world applicability of laparoscopic anatomy segmentation models

    Fiona R Kolbinger, Jiangpeng He, Jinge Ma, and Fengqing Zhu. Strategies to improve real-world applicability of laparoscopic anatomy segmentation models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2275–2284, 2024

    2024

  15. [15]

    Kolbinger, Franziska M

    Fiona R. Kolbinger, Franziska M. Rinner, Alexander C. Jenke, Matthias Carstens, Stefan Leger, et al. Anatomy segmentation in laparoscopic surgery: Comparison of machine learning and human expertise.medRxiv, 2023. doi: 10.1101/2022.11.11.22282215

    work page doi:10.1101/2022.11.11.22282215 2023

  16. [16]

    One model to use them all: training a segmentation model with com- plementary datasets.International journal of computer assisted radiology and surgery, 19(6): 1233–1241, 2024

    Alexander C Jenke, Sebastian Bodenstedt, Fiona R Kolbinger, Marius Distler, J¨ urgen Weitz, and Stefanie Speidel. One model to use them all: training a segmentation model with com- plementary datasets.International journal of computer assisted radiology and surgery, 19(6): 1233–1241, 2024

    2024

  17. [17]

    Effective disjoint representational learning for anatomical segmentation

    Priya Tomar, Aditya Parikh, Philipp Feodorovici, Jan Arensmeyer, Hanno Matthaei, Christian Bauckhage, Helen Schneider, and Rafet Sifa. Effective disjoint representational learning for anatomical segmentation. InMedical Imaging with Deep Learning, 2025

    2025

  18. [18]

    Efficient anatomy segmentation in laparoscopic surgery using multi-teacher knowledge distillation

    Lennart Maack, Finn Behrendt, Debayan Bhattacharya, Sarah Latus, and Alexander Schlae- fer. Efficient anatomy segmentation in laparoscopic surgery using multi-teacher knowledge distillation. InMedical Imaging with Deep Learning, 2024

    2024

  19. [19]

    Towards more precise automatic analysis: a systematic review of deep learning-based multi-organ seg- mentation.BioMedical Engineering OnLine, 23(1):52, 2024

    Xiaoyu Liu, Linhao Qu, Ziyue Xie, Jiayue Zhao, Yonghong Shi, and Zhijian Song. Towards more precise automatic analysis: a systematic review of deep learning-based multi-organ seg- mentation.BioMedical Engineering OnLine, 23(1):52, 2024

    2024

  20. [20]

    Transunet: Transformers make strong encoders for medical image segmentation.arXiv preprint arXiv:2102.04306, 2021

    Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation.arXiv preprint arXiv:2102.04306, 2021

    Pith/arXiv arXiv 2021

  21. [21]

    Advancements and challenges in medical image segmentation: A comprehensive survey.Sensors and AI, pages 3–29, 2025

    Guanqiu Qi, Zhiqin Zhu, Ke Li, and Han Xiao. Advancements and challenges in medical image segmentation: A comprehensive survey.Sensors and AI, pages 3–29, 2025

    2025

  22. [22]

    Improving surgical scene seman- tic segmentation through a deep learning architecture with attention to class imbalance

    Claudio Urrea, Yainet Garcia-Garcia, and John Kern. Improving surgical scene seman- tic segmentation through a deep learning architecture with attention to class imbalance. Biomedicines, 12(6):1309, 2024

    2024

  23. [23]

    Warfield, and Ali Gholipour

    Davood Karimi, Simon K. Warfield, and Ali Gholipour. Transfer learning in medical image segmentation: New insights from analysis of the dynamics of model parameters and learned representations.Artificial Intelligence in Medicine, 116:102078, 2021. ISSN 0933-3657

    2021

  24. [24]

    Curran Associates Inc., Red Hook, NY, USA, 2019

    Maithra Raghu, Chiyuan Zhang, Jon Kleinberg, and Samy Bengio.Transfusion: understanding transfer learning for medical imaging. Curran Associates Inc., Red Hook, NY, USA, 2019

    2019

  25. [25]

    Critical assessment of transfer learning for medical image segmentation with fully convolutional neural networks, 05 2020

    Davood Karimi, Simon Warfield, and Ali Gholipour. Critical assessment of transfer learning for medical image segmentation with fully convolutional neural networks, 05 2020. 17

    2020

  26. [26]

    Jumpstarting surgical computer vision

    Deepak Alapatt, Aditya Murali, Vinkle Srivastav, AI4SafeChole Consortium, Pietro Mascagni, and Nicolas Padoy. Jumpstarting surgical computer vision. InMedical Image Computing and Computer Assisted Intervention – MICCAI 2024, pages 328–338. Springer, 2024

    2024

  27. [27]

    Jaspers, Ronald L.P.D

    Tim J.M. Jaspers, Ronald L.P.D. de Jong, Yiping Li, Carolus H.J. Kusters, Franciscus H.A. Bakker, Romy C. van Jaarsveld, Gino M. Kuiper, Richard van Hillegersberg, Jelle P. Ruurda, Willem M. Brinkman, Josien P.W. Pluim, Peter H.N. de With, Marcel Breeuwer, Yasmina Al Khalil, and Fons van der Sommen. Scaling up self-supervised learning for improved surgica...

    2026

  28. [28]

    Efficient generative-adversarial u-net for multi-organ medical image segmentation.Journal of Imaging, 11(1):19, 2025

    Haoran Wang, Gengshen Wu, and Yi Liu. Efficient generative-adversarial u-net for multi-organ medical image segmentation.Journal of Imaging, 11(1):19, 2025

    2025

  29. [29]

    A unified loss for handling inter-class and intra-class imbalance in medical image segmentation

    Fei Xu, Fan Yang, Xinghui Li, and Xiaofeng Zhang. A unified loss for handling inter-class and intra-class imbalance in medical image segmentation. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 8842–8850, 2025

    2025

  30. [30]

    Endonet: a deep architecture for recognition tasks on laparoscopic videos

    Andru P Twinanda, Sherif Shehata, Didier Mutter, Jacques Marescaux, Michel De Mathelin, and Nicolas Padoy. Endonet: a deep architecture for recognition tasks on laparoscopic videos. IEEE transactions on medical imaging, 36(1):86–97, 2016

    2016

  31. [31]

    Ozan Oktay, Jo Schlemper, Lo ¨ ıc Le Folgoc, Matthew C. H. Lee, Mattias P. Heinrich, et al. Attention u-net: Learning where to look for the pancreas.CoRR, abs/1804.03999, 2018. 18

    Pith/arXiv arXiv 2018