Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Transfer Risk Map: Mitigating Pixel-level Negative Transfer in Medical Segmentation

T0 review · 4 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read A per-pixel transfer risk map can reduce negative transfer in medical segmentation fine-tuning, with gains of 4.37% on FeTS 2021 and 1.81% on iSeg-2019.

desk verdict Incremental pixel-weighted fine-tuning that mostly works in tables but lacks significance testing and ignores a spatial-alignment issue. read the letter →

arxiv 2502.02340 v1 pith:GSDNLCOO submitted 2025-02-04 cs.CV

Shutong Duan , Jingyun Yang , Yang Tan , Guoqing Zhang , Yang Li , Xiao-Ping Zhang This is my paper · ORCID
classification cs.CV
keywords medicalimagesegmentationtransferlearningnegativetransferabilityestimationpixel-levelweightingfine-tuningbrainMRIfew-shot
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to show that negative transfer — the situation where knowledge from a pretrained source task hurts performance on the target task — is not uniform across an image. The authors compute a transfer risk map from pixel-wise LEEP transferability scores, reweight the per-pixel segmentation loss with this map, and average the loss over foreground pixels only. If the claim holds, transfer learning between different MRI modalities and tasks becomes safer without needing access to source data or a new architecture. The reported gains are 4.37% Dice on the FeTS 2021 brain tumor dataset and 1.81% on the iSeg-2019 brain tissue dataset, with an additional 2.9% average gain in a few-shot setting.

What carries the argument

The load-bearing object is the transfer risk map $w$, a fixed spatial weight matrix computed once from pixel-level LEEP scores on the target dataset. Each entry $w_{j,k} = 10^{t^s_{j,k}}$ turns a min-max normalized LEEP hardness score into an upweighting factor for the loss at coordinate $(j,k)$, and the exponential base 10 makes the emphasis on hard pixels nonlinear. The map is applied to every training image at the same coordinates, and the reweighted per-pixel loss is averaged over foreground pixels to counter class imbalance. This one map is what carries the claimed effect: it concentrates fine-tuning pressure on the regions where transfer is riskiest.

What would settle it

Run the same transfer fine-tuning on a brain MRI dataset whose slices from different subjects are randomly rotated, shifted, or resampled before training, keeping the labels in the same coordinate frame; if the reported Dice gains disappear or reverse, the central claim depends on the unstated alignment assumption.

Watch

Extended reading notes

Core claim

The central discovery claimed is that pixel-level transferability, measured by adapting LEEP (Log Expected Empirical Prediction), a transferability score for classification, to dense prediction, predicts where negative transfer will hurt during fine-tuning, and that weighting the loss at exactly those pixels improves target performance. Concretely, the paper builds a single transfer risk map $w$ over image coordinates from min-max normalized pixel-level LEEP scores, scaled by $10^{(\cdot)}$ to emphasize high-risk pixels, then fine-tunes a frozen-encoder U-Net with the loss $\sum w_{j,k}\,\ell_{j,k}$ averaged over foreground pixels only. Across cross-modality and cross-task brain MRI experiments, the weighted fine-tuning beats vanilla fine-tuning, class-weighted fine-tuning, and a prior transferability-map baseline, with average Dice gains of 4.37% on FeTS 2021 and 1.81% on iSeg-2019, and a 2.9% gain with as few as 50 target slices.

Load-bearing premise

The fixed per-coordinate transfer risk map assumes that the 2D slices used for training are spatially aligned across subjects, because the same pixel coordinate is treated as the same anatomical location for every image; the paper describes no registration or alignment step.

Editorial extensions

If this is right

  • If the central claim is correct, cross-modality and cross-task transfer in medical segmentation can be improved by reweighting the loss alone, with no access to source data.
  • The weighting scheme should remain useful in few-shot regimes, since the paper reports a 2.9% average Dice improvement even when the target has only 50 labeled slices.
  • Because the method lives entirely in the loss, it can be combined with any segmentation architecture, including the vision transformer extension the authors mention.
  • Averaging the reweighted loss over foreground pixels only prevents a large well-learned background from diluting the loss, which matters for small lesions and fine structures.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves implicit that its fixed coordinate map presumes anatomical alignment across subjects; a natural extension would be to compute per-image risk maps or add a registration step before building the map.
  • A straightforward testable follow-up is to update the transfer risk map periodically during fine-tuning rather than once at the start, since the paper only compares maps before and after training.
  • The same recipe could apply to other dense prediction tasks, such as depth estimation or panoptic segmentation, wherever a pixel-wise transferability score can be defined, though the paper does not make this claim.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes a transferability-guided weighting scheme for fine-tuning medical image segmentation models. A pixel-level transfer risk map is computed using a segmentation-adapted LEEP score evaluated on the target training data, and this map is used to reweight the per-pixel loss during fine-tuning, with normalization by foreground size to mitigate class imbalance. Experiments on two brain MRI datasets (FeTS 2021 and iSeg-2019) report consistent Dice improvements over vanilla fine-tuning, class-weighted fine-tuning, and a transferability-map baseline, along with improvements in a few-shot setting. The central claim is that the method avoids negative transfer across modalities and tasks.

Significance. If substantiated, the proposed method would be a simple and practical addition to the transfer-learning toolbox for medical segmentation: it requires only a pre-trained source model and labeled target data, and it operates at the pixel level where negative transfer is plausibly non-uniform. The use of LEEP on a per-pixel basis is a natural extension of existing transferability metrics, and the reported gains are consistent across many transfer pairs. However, the evidence currently falls short of the paper's own claims: there is no from-scratch baseline, no statistical significance testing, and a load-bearing spatial-alignment assumption is never stated or verified. The method's utility as a mechanism for avoiding negative transfer therefore remains unproven, although the core idea is defensible and testable.

major comments (4)
  1. [Section III-C, Tables I and III; Abstract/Conclusion] The central claim that the method 'indeed avoids negative transfer' (Abstract, line 6 and Conclusion) is not supported by the reported experiments. Negative transfer is defined relative to training without transfer (i.e., from scratch), yet all comparisons in Tables I, II, and III are between fine-tuning variants. A from-scratch baseline trained on the target data is necessary to establish whether the proposed weighting actually prevents the source model from hurting target performance, as opposed to merely improving a fine-tuning procedure.
  2. [Section II-C, Equations (4)-(7)] The transfer risk map w is a single W x H mask computed once over all target images at fixed coordinates (j,k) and then applied to every training image in Eq. (7). This assumes that pixel coordinate (j,k) corresponds to the same anatomical structure across subjects. No registration, atlas resampling, or slice-selection procedure is described in Section III-B, and for unaligned brain MRI volumes this assumption is unlikely to hold. Without an alignment step or a demonstration that the mask is anatomically consistent, the weighting may emphasize different structures in different images, and the reported gains cannot be confidently attributed to pixel-level transfer risk.
  3. [Section III-B and Tables I-III] The paper provides no error bars, significance tests, or a description of the train/test split. It is not stated how the target data are divided (e.g., patient-level split, which 2D slices are used for training vs. testing, validation procedure for early stopping), nor how many random seeds the reported Dice scores are averaged over. Given that the text repeatedly describes improvements as 'significantly enhanced' (e.g., Section III-C), the absence of any statistical support is a serious gap that prevents evaluation of whether the observed gains are reliable.
  4. [Section II-B, Equations (2)-(4)] The pixel-level LEEP computation is underspecified. The text says the global feature vector is decomposed into pixel-wise features, but it does not state which network layer is used, how feature resolution is matched to the input resolution (given the U-Net's downsampling), or how the empirical conditional distribution P(y_{j,k} | z) is estimated per pixel. Without these details, the transfer risk map in Eq. (4) is not reproducible, and it is unclear whether the map truly reflects per-pixel transferability as opposed to some artifact of feature-map interpolation.
minor comments (4)
  1. [Section II-C, Eq. (6)] The exponential scaling base 10 is introduced without any ablation or justification; the choice of base strongly affects the dynamic range of the weights, and a sensitivity analysis or rationale should be provided.
  2. [Section III-D, Table III] The text states 'a 2.4% improvement and a 2.9% gain in Dice score in average' for the 50-sample case, but the table shows a relative improvement of approximately 2.85% for that row; the wording is inconsistent and should be clarified.
  3. [Section II-C, Eq. (7)] The loss is averaged over foreground pixels only, but the numerator sums over all pixels; this means background pixels contribute to the loss value in the numerator but not to the denominator. The authors should clarify whether this is intentional and discuss the effect on gradients, since the background loss is still backpropagated despite being excluded from the normalization.
  4. [Section III-C, Tables I and II] The reported 'overall gain' percentages appear to be relative improvements over vanilla fine-tuning (e.g., 4.37% for FeTS 2021), but this is not stated explicitly; the text should define how these average gains are computed.

Circularity Check

1 steps flagged · score 4.0 of 10

The central Dice-gain result is not circular, but the before/after transfer-risk validation in Sec. III-E reduces to a supervised training-progress measure because LEEP is computed and refit on the same target labels.

  1. fitted input called prediction [Section III-E ('Effectiveness of Negative Transfer Mitigation'), Fig. 2; Eqs. (2)-(6)]
    "It is evident that our fine-tuning approach has effectively condensed the extensive regions of high transfer risk into well-defined, smaller segments, concurrently reducing the transfer hardness."

    The risk map w comes from Eq. (4): t_{j,k}=LEEP(θ, {x^{j,k}_i, y^{j,k}_i}) on the target set D, and LEEP in Eq. (2) builds P̂(y|z) from those same target labels. After fine-tuning on D, θ's likelihood on D rises, so LEEP-based 'transfer hardness' falls at least partly because the metric is refit to the labels used in training, not because an independent negative-transfer effect was removed. The before/after visualization is thus not an independent validation; it is a supervised training-progress measure presented as evidence of negative-transfer mitigation.

full rationale

The main quantitative claims (4.37% Dice gain on FeTS2021, 1.81% on iSeg-2019, 2.9% few-shot gain) compare fine-tuning variants on public benchmarks and do not reduce to the risk-map construction by definition; the risk map is a fixed supervised weight computed before training, not a prediction of the final Dice. The method's dependence on LEEP is externally grounded in [18], so the self-citation [22] (authors' own prior work) is used for comparison and for choosing LEEP but is not the load-bearing derivation. However, Section III-E presents the pre/post fine-tuning transfer-risk maps as validation of negative-transfer mitigation; because LEEP uses the empirical conditional distribution estimated on the same target labels that are used to fine-tune the model, the observed decrease in 'transfer hardness' is largely a consequence of the metric being fitted to the training labels. This is a partial circularity in the supporting argument, though the independent benchmark results keep the central claim from being fully circular. The lack of an explicit train/test split and the unstated assumption that fixed pixel coordinates correspond to the same anatomy across slices are validity risks, not circularity.

Assumptions & free parameters 1 free parameters · 4 assumptions · 1 invented entities

The central method depends on the untested assumption that pixel-level LEEP scores quantify transfer risk, on a spatial-alignment assumption for the risk map, and on several heuristic choices (base-10 exponent, foreground-only normalization). The only hand-chosen numerical constant explicitly identified is the exponential base 10. No fitted physical parameters are involved.

free parameters (1)
  • Exponential scaling base = 10
    Section II-C, Eq. (6): the transfer risk map applies 10 raised to the normalized transferability. The base is hand-chosen and not ablated, and it controls how strongly high-risk pixels are emphasized.
assumptions (4)
  • domain assumption LEEP score computed on target data with source model predictions is a valid per-pixel measure of negative transfer risk.
    Section II-B/C: the entire method rests on treating pixel-level LEEP as a proxy for transfer hardness; no evidence is given that this proxy corresponds to actual negative transfer.
  • ad hoc to paper Pixel coordinates are spatially aligned across target images so a single risk map can be applied to every image.
    Section II-C, Eq. (4) defines t_{j,k} at fixed coordinates across the dataset, and Eq. (7) applies the same w_{j,k} to all images; no registration or alignment is described.
  • domain assumption The source-free setting permits estimating LEEP's empirical conditional distribution from target data only.
    Section II-A/B: the paper uses LEEP with source model predictions on target data, implicitly assuming this estimate is sufficient for transferability; this is standard LEEP usage but unverified here.
  • domain assumption Foreground-only normalization of the loss is a valid class-imbalance correction and does not distort the training objective.
    Section II-C, Eq. (7): the denominator counts only foreground pixels; this choice is motivated heuristically and not ablated.
invented entities (1)
  • Transfer risk map
    purpose: Spatial weighting mask for the fine-tuning loss to emphasize pixels with high estimated transfer risk.
    Defined in Section II-C, Eqs. (5)-(6) from LEEP scores; the paper provides no external validation that the map corresponds to genuine per-pixel negative transfer risk, so it is an unverified construct.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Transfer Risk Map: Mitigating Pixel-level Negative Transfer in Medical Segmentation." pith.science (2026). https://pith.science/paper/GSDNLCOO

@misc{pith2026250202340,
  author       = {Pith},
  title        = {Pith review of: Transfer Risk Map: Mitigating Pixel-level Negative Transfer in Medical Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GSDNLCOO}},
  note         = {Machine review of arXiv:2502.02340}
}
read the original abstract

How to mitigate negative transfer in transfer learning is a long-standing and challenging issue, especially in the application of medical image segmentation. Existing methods for reducing negative transfer focus on classification or regression tasks, ignoring the non-uniform negative transfer risk in different image regions. In this work, we propose a simple yet effective weighted fine-tuning method that directs the model's attention towards regions with significant transfer risk for medical semantic segmentation. Specifically, we compute a transferability-guided transfer risk map to quantify the transfer hardness for each pixel and the potential risks of negative transfer. During the fine-tuning phase, we introduce a map-weighted loss function, normalized with image foreground size to counter class imbalance. Extensive experiments on brain segmentation datasets show our method significantly improves the target task performance, with gains of 4.37% on FeTS2021 and 1.81% on iSeg2019, avoiding negative transfer across modalities and tasks. Meanwhile, a 2.9% gain under a few-shot scenario validates the robustness of our approach.

Figures

Figures reproduced from arXiv: 2502.02340 by the authors.

Figure 1
Figure 1. Illustration of the proposed transfer risk map guided [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The pixel-level transfer risk maps before and after [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Understanding Knowledge Transferability for Transfer Learning: A Survey

    cs.LG 2025-07 conditional novelty 4.0 of 10

    A survey that classifies transferability metrics by knowledge modality (dataset vs. model) and granularity (task vs. instance), with a theoretical primer and applications to eight learning paradigms.

Reference graph

Works this paper leans on

29 extracted references · 21 canonical work pages · cited by 1 Pith paper

  1. [22]

    Efficient prediction of model transferability in semantic segmentation tasks,

    Y . Tan, Y . Li, Y . Li, and X.-P. Zhang, “Efficient prediction of model transferability in semantic segmentation tasks,” in 2023 IEEE Interna- tional Conference on Image Processing (ICIP) , 2023, pp. 720–724

    work page 2023

  2. [1]

    Medical image segmentation using deep learning: A survey,

    R. Wang, T. Lei, R. Cui, B. Zhang, H. Meng, and A. K. Nandi, “Medical image segmentation using deep learning: A survey,”IET image processing, vol. 16, no. 5, pp. 1243–1267, 2022

    work page 2022

  3. [2]

    Deep learning tech- niques for medical image segmentation: achievements and challenges,

    M. H. Hesamian, W. Jia, X. He, and P. Kennedy, “Deep learning tech- niques for medical image segmentation: achievements and challenges,” Journal of digital imaging , vol. 32, pp. 582–596, 2019

    work page 2019

  4. [3]

    Transfer learning techniques for medical image analysis: A review,

    P. Kora, C. P. Ooi, O. Faust, U. Raghavendra, A. Gudigar, W. Y . Chan, K. Meenakshi, K. Swaraja, P. Plawiak, and U. R. Acharya, “Transfer learning techniques for medical image analysis: A review,” Biocybernetics and Biomedical Engineering , vol. 42, no. 1, pp. 79–107, 2022

    work page 2022

  5. [4]

    Transfer learning in med- ical image segmentation: New insights from analysis of the dynamics of model parameters and learned representations,

    D. Karimi, S. K. Warfield, and A. Gholipour, “Transfer learning in med- ical image segmentation: New insights from analysis of the dynamics of model parameters and learned representations,” Artificial intelligence in medicine, vol. 116, p. 102078, 2021

    work page 2021

  6. [5]

    A survey on transfer learning,

    S. J. Pan and Q. Yang, “A survey on transfer learning,” IEEE Trans- actions on knowledge and data engineering , vol. 22, no. 10, pp. 1345– 1359, 2009

    2009

  7. [6]

    Characterizing and avoiding negative transfer,

    Z. Wang, Z. Dai, B. Poczos, and J. Carbonell, “Characterizing and avoiding negative transfer,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , June 2019

    work page 2019

  8. [7]

    Domain adaptation for medical image analysis: a survey,

    H. Guan and M. Liu, “Domain adaptation for medical image analysis: a survey,” IEEE Transactions on Biomedical Engineering , vol. 69, no. 3, pp. 1173–1185, 2021

    2021

Show all 29 references
  1. [8]

    A survey on negative transfer,

    W. Zhang, L. Deng, L. Zhang, and D. Wu, “A survey on negative transfer,” IEEE/CAA Journal of Automatica Sinica , vol. 10, no. 2, pp. 305–329, 2023

    2023

  2. [9]

    Finding the most trans- ferable tasks for brain image segmentation,

    Y . Li, Y . Tan, J. Yang, Y . Li, and X.-P. Zhang, “Finding the most trans- ferable tasks for brain image segmentation,” in 2022 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) . IEEE, 2022, pp. 1620–1625

    2022

  3. [10]

    A cross-dataset adaptive domain selection transfer learning framework for motor imagery-based brain-computer interfaces,

    J. Jin, G. Bai, R. Xu, K. Qin, H. Sun, X. Wang, and A. Cichocki, “A cross-dataset adaptive domain selection transfer learning framework for motor imagery-based brain-computer interfaces,” Journal of Neural Engineering, vol. 21, no. 3, p. 036057, 2024

    2024

  4. [11]

    Non-negative transfer learning with consistent inter-domain distribution,

    Z. Peng, Y . Jia, and J. Hou, “Non-negative transfer learning with consistent inter-domain distribution,” IEEE Signal Processing Letters , vol. 27, pp. 1720–1724, 2020

    2020

  5. [12]

    Dual transfer learning,

    M. Long, J. Wang, G. Ding, W. Cheng, X. Zhang, and W. Wang, “Dual transfer learning,” in Proceedings of the 2012 SIAM International Conference on Data Mining . SIAM, 2012, pp. 540–551

    2012

  6. [13]

    A kernel two-sample test,

    A. Gretton, K. M. Borgwardt, M. J. Rasch, B. Sch ¨olkopf, and A. Smola, “A kernel two-sample test,” The Journal of Machine Learning Research, vol. 13, no. 1, pp. 723–773, 2012

    2012

  7. [14]

    On information and sufficiency,

    S. Kullback and R. A. Leibler, “On information and sufficiency,” The annals of mathematical statistics , vol. 22, no. 1, pp. 79–86, 1951

    1951

  8. [15]

    Improving eeg-based emotion classification using conditional transfer learning,

    Y .-P. Lin and T.-P. Jung, “Improving eeg-based emotion classification using conditional transfer learning,” Frontiers in human neuroscience , vol. 11, p. 334, 2017

    2017

  9. [16]

    Transferability and hardness of supervised classification tasks,

    A. T. Tran, C. V . Nguyen, and T. Hassner, “Transferability and hardness of supervised classification tasks,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2019, pp. 1395–1405

    2019

  10. [17]

    An information-theoretic approach to transferability in task transfer learning,

    Y . Bao, Y . Li, S.-L. Huang, L. Zhang, L. Zheng, A. Zamir, and L. Guibas, “An information-theoretic approach to transferability in task transfer learning,” in 2019 IEEE international conference on image processing (ICIP). IEEE, 2019, pp. 2309–2313

    2019

  11. [18]

    Leep: A new measure to evaluate transferability of learned representations,

    C. Nguyen, T. Hassner, M. Seeger, and C. Archambeau, “Leep: A new measure to evaluate transferability of learned representations,” in International Conference on Machine Learning . PMLR, 2020, pp. 7294–7305

    2020

  12. [19]

    Logme: Practical assessment of pre-trained models for transfer learning,

    K. You, Y . Liu, J. Wang, and M. Long, “Logme: Practical assessment of pre-trained models for transfer learning,” in International Conference on Machine Learning . PMLR, 2021, pp. 12 133–12 143

    2021

  13. [20]

    Otce: A transferability metric for cross- domain cross-task representations,

    Y . Tan, Y . Li, and S.-L. Huang, “Otce: A transferability metric for cross- domain cross-task representations,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 15 779–15 788

    2021

  14. [21]

    Where and how to transfer: Knowledge aggregation-induced transferability perception for unsupervised domain adaptation,

    J. Dong, Y . Cong, G. Sun, Z. Fang, and Z. Ding, “Where and how to transfer: Knowledge aggregation-induced transferability perception for unsupervised domain adaptation,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 3, pp. 1664–1681, 2024

    2024

  15. [23]

    Deep semantic segmentation of natural and medical images: a review,

    S. Asgari Taghanaki, K. Abhishek, J. P. Cohen, J. Cohen-Adad, and G. Hamarneh, “Deep semantic segmentation of natural and medical images: a review,” Artificial Intelligence Review , vol. 54, pp. 137–178, 2021

    2021

  16. [24]

    The federated tumor segmentation (fets) challenge,

    S. Pati, U. Baid, M. Zenk, B. Edwards, M. Sheller, G. A. Reina, P. Foley, A. Gruzdev, J. Martin, S. Albarqouni et al. , “The federated tumor segmentation (fets) challenge,” arXiv preprint arXiv:2105.05874, 2021

    2021 arXiv

  17. [25]

    Openfl: An open-source framework for federated learning,

    G. A. Reina, A. Gruzdev, P. Foley, O. Perepelkina, M. Sharma, I. Davidyuk, I. Trushkin, M. Radionov, A. Mokrov, D. Agapov et al. , “Openfl: An open-source framework for federated learning,” arXiv preprint arXiv:2105.06413, 2021

    2021 arXiv

  18. [26]

    Advancing the cancer genome atlas glioma mri collections with expert segmentation labels and radiomic features,

    S. Bakas, H. Akbari, A. Sotiras, M. Bilello, M. Rozycki, J. S. Kirby, J. B. Freymann, K. Farahani, and C. Davatzikos, “Advancing the cancer genome atlas glioma mri collections with expert segmentation labels and radiomic features,” Scientific data, vol. 4, no. 1, pp. 1–13, 2017

    2017

  19. [27]

    Multi-site infant brain segmentation algorithms: the iseg-2019 challenge,

    Y . Sun, K. Gao, Z. Wu, G. Li, X. Zong, Z. Lei, Y . Wei, J. Ma, X. Yang, X. Feng et al. , “Multi-site infant brain segmentation algorithms: the iseg-2019 challenge,” IEEE Transactions on Medical Imaging , vol. 40, no. 5, pp. 1363–1376, 2021

    2019

  20. [28]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015 , N. Navab, J. Horneg- ger, W. M. Wells, and A. F. Frangi, Eds. Cham: Springer International ...

    2015

  21. [29]

    Transfer learning with class-weighted and focal loss function for automatic skin cancer classification,

    D. N. Le, H. X. Le, L. T. Ngo, and H. T. Ngo, “Transfer learning with class-weighted and focal loss function for automatic skin cancer classification,” arXiv preprint arXiv:2009.05977 , 2020

    2009 arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.