ActiveSSF: An Active-Learning-Guided Self-Supervised Framework for Long-Tailed Megakaryocyte Classification

Gege Yuan; Linghao Zhuang; Xingyue Zhao; Ying Zhang; Zhiping Jiang

REVIEW 4 major objections 4 minor 15 references

ActiveSSF: An Active-Learning-Guided Self-Supervised Framework for Long-Tailed Megakaryocyte Classification

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

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

Pith's one-line read ActiveSSF is a two-stage active-learning wrapper for self-supervised pretraining that improves megakaryocyte classification, with the largest gains on rare subtypes.

desk verdict Solid applied medical-imaging paper with consistent, statistically significant gains, but the active-selection mechanism is unverified and the method description has fixable inconsistencies. read the letter →

arxiv 2502.08200 v2 pith:XDCZNJDJ submitted 2025-02-12 cs.CV

Linghao Zhuang , Ying Zhang , Gege Yuan , Xingyue Zhao , Zhiping Jiang This is my paper · ORCID

classification cs.CV

keywords activelearningself-supervisedlong-taileddistributionmegakaryocyteclassificationmedicalimageanalysisprototypeclusteringsampleselectionHSVcolorspace

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 tries to establish that self-supervised pretraining for medical cell classification can be materially improved by choosing which unlabeled images to train on, rather than using all of them. It introduces ActiveSSF, a two-stage pipeline that cleans stained-slide images, builds prototypes from labeled examples, and then admits unlabeled examples into pretraining with thresholds that favor under-represented subtypes. A sympathetic reader would care because megakaryocyte subtypes are long-tailed and hard to distinguish, and the reported gains are concentrated on the rare subtypes that matter for diagnosing myelodysplastic syndromes. The strongest evidence is the comparison on the authors' 11-subtype dataset: ConvMAE with ActiveSSF reaches 81.10% accuracy versus 76.66% without it.

What carries the argument

The load-bearing mechanism is a two-stage active-selection loop. First, Cell Region Filtering crops informative cell regions from noisy slides using Gaussian blur, two-stage K-means color clustering, and HSV thresholds set with clinical priors. Second, Active Sample Selection embeds labeled samples with a ResNet, clusters them into 11 prototypes via K-means, and then accepts an unlabeled region into the pretraining pool only if its distance to the nearest prototype is below a density-aware threshold; the threshold is lower for small clusters, so rare subtypes are admitted more liberally. Pretraining with masked autoencoding is then run on this selected pool rather than on all unlabeled images.

What would settle it

Measure the per-subtype composition of the unlabeled samples that ActiveSSF actually selects. If the selected pool does not contain a higher fraction of rare-subtype examples than random selection (as judged by expert labels on a holdout subset), or if replacing the prototype-distance selection with random selection of the same size does not change downstream accuracy, the central claim that active selection drives the gains would be falsified.

Watch

Extended reading notes

Core claim

The paper claims that guiding self-supervised pretraining with an active-learning selection step—rather than consuming all unlabeled data blindly—substantially improves fine-grained classification of megakaryocyte subtypes in stained bone marrow slides. On the authors' 11-subtype clinical dataset, wrapping the ConvMAE self-supervised model in ActiveSSF raises accuracy from 76.66% to 81.10%, and the largest per-subtype gains are in rare classes such as HL-MK (PR-AUC from 72.24% to 81.12%) and CA-MK (68.09% to 73.67%). The same pattern holds across all four self-supervised baselines tested, with statistically significant improvements in accuracy, PR-AUC, and F1-score.

Load-bearing premise

The method assumes that K-means clustering of ResNet features from the 7,577 labeled images produces 11 clusters that line up with the 11 biological subtypes, so that distance to the nearest cluster center is a reliable way to pick informative unlabeled images; if rare-subtype prototypes are noisy, selection could reinforce the original imbalance rather than correct it.

Editorial extensions

If this is right

With ConvMAE, ActiveSSF raises accuracy from 76.66% to 81.10% and PR-AUC from 79.72% to 84.37%.
Rare subtypes benefit most: for example, HL-MK PR-AUC rises from 72.24% to 81.12% and CA-MK from 68.09% to 73.67% when ConvMAE is wrapped in ActiveSSF.
The improvement is consistent across MoCo v3, ConvMAE, MIXMIM, and MAE, with all reported paired t-tests below 0.05.
Ablations on MAE show that both Cell Region Filtering and Active Sample Selection contribute incrementally, in that order.
The method turns 10,000 unlabeled slide images into a curated pretraining set that improves an 11-class clinical classification task.

Reading between the lines

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

If the mechanism is right, the same two-stage selection could be dropped into other long-tailed medical image tasks, but only if prototype clusters align with clinically meaningful classes; the paper does not verify this alignment directly.
A direct comparison against random selection of the same number of unlabeled images would isolate whether the benefit comes from active selection or simply from more pretraining data; the paper's ablations do not include this control.
The density-aware threshold formula assumes cluster size tracks class frequency; for subtypes with high morphological variation, which may split into several clusters, the selection could over- or under-sample relative to the true class balance.

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.

Desk Editor's Note

read the letter

Quick take: this is a decent applied paper with real, consistent gains across several SSL backbones, but the long-tail rebalancing story is asserted, not demonstrated. The K-means prototypes are never shown to correspond to the 11 subtypes, and there are internal inconsistencies in the method description that need cleaning up.

What's genuinely new: it is the first deep-learning fine-grained classification of megakaryocyte subtypes, on a private clinical dataset, and the idea of using active sample selection to guide self-supervised pretraining is a sensible application to a long-tailed medical task. The experiments are the strength: gains over baselines are consistent across MoCo v3, ConvMAE, MIXMIM, and MAE, with paired t-tests and five-fold cross-validation. The ablation shows each module adds something, though the increments are modest.

The soft spots are real but not fatal. The main one: K-means with K=11 on ResNet features of 7,577 labeled samples is run without any check that clusters are pure subtypes. In a long-tailed distribution, K-means can split common classes or merge rare ones, so cluster size is not a reliable proxy for subtype rarity. If clusters don't align, the dynamic threshold in Eq. (8) is lowering thresholds for large or artifact clusters, not for rare classes. The paper reports no per-cluster or per-subtype selection counts, so we can't tell whether the selected unlabeled set is actually balanced. That needs fixing, at minimum with prototype-purity numbers and selection counts per subtype.

Second, the method description has a genuine inconsistency: Algorithm 1 states alpha=0.5, but the implementation text says alpha=1 and the dynamic threshold was "17±2," which doesn't follow from Eq. (8) as written. This makes it hard to know what was actually run. Also, the HSV thresholds, fill rate, and bounding box size are all tuned on the same dataset without a held-out validation split; the authors should discuss sensitivity.

Less important: no comparison to standard long-tail techniques (reweighting, class-balanced sampling) or to a random selection baseline, so the active-selection component isn't isolated. No code or data release, which limits reproducibility.

Who this is for: people working on medical cell classification and SSL practitioners looking for an applied recipe. It deserves a serious referee; I'd send it out, but with a clear request for prototype validation, consistency fixes, and one or two stronger baselines. The central empirical claim—that this pipeline helps rare-class accuracy—is plausible and supported, so it's not a reject on substance.

Referee Report

4 major / 4 minor

Summary. The paper proposes ActiveSSF, a two-stage active-learning-guided self-supervised framework for classifying eleven megakaryocyte subtypes from stained bone-marrow slides. Stage 1 (Cell Region Filtering) suppresses background via Gaussian blur, two-stage K-means clustering, HSV thresholding, and morphological filtering; Stage 2 (Active Sample Selection) clusters ResNet50 features of the labeled data into 11 prototypes and uses a density-aware threshold (Eq. (8)) to select unlabeled samples for self-supervised pretraining. Experiments on a clinical dataset of 7,577 labeled and 10,000 unlabeled images report accuracy/PR-AUC/F1 gains across four self-supervised backbones (MoCo v3, ConvMAE, MIXMIM, MAE), with paired t-tests and an ablation study. The central claim is that ActiveSSF improves overall accuracy and especially rare-subtype recognition by rebalancing the pretraining data through adaptive sample selection.

Significance. If the mechanism is verified, ActiveSSF addresses a practically important problem: making self-supervised pretraining effective for fine-grained, long-tailed medical image classification with very limited labeled data. The paper's strengths include consistent gains across multiple self-supervised backbones, statistical testing with paired t-tests, an ablation isolating the contributions of cell-region filtering and active sample selection, and a clinically curated dataset with expert annotations. However, the paper as written does not establish the load-bearing causal link between its adaptive thresholding mechanism and the reported rare-subtype improvements: the prototype-to-subtype alignment is unverified, Eq. (8) appears to behave opposite to the prose, and the implementation details of the threshold are internally inconsistent. These issues must be resolved before the central claim can be accepted.

major comments (4)

[§II.C.2, Eq. (8)] The text states that the mechanism 'automatically lowers the threshold for rare subtypes,' but Eq. (8) as written has the opposite behavior: for a rare cluster (ni ≪ nmax), the factor (1 − ni/nmax)^α is close to 1, driving thresholdi toward UB, i.e., a more permissive threshold. If the intended behavior is to admit more rare samples, the prose should say 'raises'; if the intended behavior is genuinely 'lowers,' the formula must be revised. Additionally, LB and UB are defined only as 'minimum and maximum distances within each cluster,' which does not specify whether they are computed per cluster or globally, making Eq. (8) non-reproducible.
[§II.C.1 and Eq. (9)] The active-selection mechanism assumes that the K=11 K-means clusters on ResNet50 features correspond to the 11 biological subtypes, so that cluster size ni is a proxy for subtype rarity. This assumption is not verified. Under a long-tailed distribution, K-means can split common classes and merge rare classes, or group image artifacts rather than morphologically meaningful subtypes. Please report a cluster-to-subtype contingency table on the 7,577 labeled samples and/or per-subtype selection counts from the 7,299 filtered unlabeled samples. Without such evidence, the rare-subtype gains in Table I cannot be attributed to the density-aware thresholding mechanism.
[§III.B and Algorithm 1] The implementation paragraph states that the dynamic threshold used in sample selection was '17 ± 2' with α=1, while Algorithm 1 sets α=0.5 and Eq. (8) produces a per-cluster threshold from LB and UB, not a single global value. This inconsistency makes it impossible to reconstruct the actual selection procedure. Please specify the exact formula used, the values of LB, UB, and α, and whether a single global threshold or per-cluster thresholds were applied.
[§II.B and §II.C] Several preprocessing and selection hyperparameters—HSV hue/saturation ranges, fill-rate 70%, minimum bounding-box size 70×70, Gaussian σ=1.5, scaling factor α, and the distance bounds LB/UB—appear to be selected on the same dataset without a held-out validation split. Since the core contribution is a data-selection mechanism, the reported gains should be shown to be robust to these choices, for example through a sensitivity analysis or nested cross-validation.

minor comments (4)

[§III.C] The phrase 'state-of-the-art performance' is stronger than what the experiments support, since Table II compares ActiveSSF only with the same self-supervised baselines with and without the framework, not with other active-learning or long-tail rebalancing methods; please temper this claim or add such comparisons.
[§II.B.1] The two-stage K-means clustering (20 clusters then 10 clusters) is described only in prose; Algorithm 1 does not list the cluster counts, so please state K1=20 and K2=10 explicitly in the algorithm.
[Table I] Per-class PR-AUC for rare classes shows large standard deviations (e.g., ViT HL-MK 43.85 ± 16.73), but no per-class significance tests are reported; adding them would strengthen the rare-subtype claim.
[§III.B] The filtering pipeline produces 7,299 samples from 10,000 slides; please state the number of images rejected or accepted at each substage (Gaussian/K-means, HSV mask, morphological filtering) to clarify the dataset construction.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ActiveSSF is an empirical pipeline whose selection mechanism is not equivalent to its evaluation by construction.

full rationale

ActiveSSF is an empirical framework rather than a mathematical derivation, and its central claim—improved rare-subtype classification after active sample selection for self-supervised pretraining—is tested against held-out folds in Table II and Table I. The prototypes in Section II.C.1 are generated by K-means on labeled ResNet50 features, and Eq. (8)-(9) use cluster sizes to set selection thresholds; this is a heuristic design choice, not a quantity that is defined in terms of the final accuracy or PR-AUC it is claimed to predict. The final fine-tuned classifier is trained on labeled data, and the reported gains come from comparing the same self-supervised backbone with and without ActiveSSF, so no equation or fitted parameter reduces to the reported outcome by construction. The paper does not invoke a uniqueness theorem, does not import a load-bearing ansatz from the authors' prior work, and does not rename a known result as a new one. The implementation inconsistency (Algorithm 1 sets α=0.5 while the implementation paragraph says α=1 and threshold '17 ± 2') is a reproducibility concern, not circularity. The unverified assumption that the 11 K-means clusters align with the 11 subtypes is an empirical risk that could undermine the mechanism, but it is not a logical circularity. Minor co-authored references [6], [7] appear in the introduction and are not load-bearing for the method's validity. Therefore the derivation chain is self-contained enough that no circular step can be exhibited.

Assumptions & free parameters 9 free parameters · 4 assumptions · 0 invented entities

All reported improvements rest on a chain of manual hyperparameters and domain assumptions: the preprocessing thresholds, prototype counts, and selection thresholds are set by the authors on the same dataset, and the alignment of K-means clusters with biological subtypes is assumed, not verified.

free parameters (9)

Gaussian sigma = 1.5
Chosen in Algorithm 1 without sensitivity analysis.
K-means clusters (preprocessing) = 20 then 10
Two-stage clustering counts chosen by hand.
Number of prototypes = 11
Set equal to the number of subtypes; may not reflect actual feature clusters.
HSV purple mask thresholds = H 30-140, S 100-255, V 0-255
Determined empirically with clinical experts; not validated on held-out data.
HSV deep blue mask thresholds = H 95-105, S 150-255, V 50-255
Described as 'e.g.', so exact values are not fixed.
Fill rate threshold = 0.7
Manual choice in Algorithm 1.
Minimum bounding box size = 70x70 pixels
Manual choice with no ablation.
Scaling factor alpha = 0.5 or 1 (inconsistent)
Algorithm 1 says 0.5; Implementation Details says 1.
Distance bounds LB, UB = LB, UB with report 17 +/- 2
The text says 'dynamic threshold set to 17 +/- 2', but Eq. (8) defines thresholds from per-cluster LB/UB, so the exact value is ambiguous.

assumptions (4)

domain assumption ResNet50 features capture morphological discriminability of megakaryocyte subtypes
Section II-C1 uses ResNet features for clustering without validating that the feature space separates the 11 classes.
domain assumption K-means with K=11 on labeled features yields clusters that correspond to the 11 biological subtypes
Section II-C1 assumes prototype-to-class alignment; no label-cluster purity is reported.
domain assumption HSV thresholds derived from clinical priors correctly isolate megakaryocyte regions across all unlabeled slides
Section II-B2 states thresholds were 'carefully determined' with experts but does not assess generalization across the 10,000 images.
domain assumption Self-supervised pretraining on the actively selected unlabeled set transfers to the downstream fine-tuning task better than training on the full unlabeled set
The experimental design compares with/without selection, but the mechanism of transfer is not analyzed; this is the central empirical premise.

how reviews work

0 comments

Cite this review

Pith. "Pith review of ActiveSSF: An Active-Learning-Guided Self-Supervised Framework for Long-Tailed Megakaryocyte Classification." pith.science (2026). https://pith.science/paper/XDCZNJDJ

@misc{pith2026250208200,
  author       = {Pith},
  title        = {Pith review of: ActiveSSF: An Active-Learning-Guided Self-Supervised Framework for Long-Tailed Megakaryocyte Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XDCZNJDJ}},
  note         = {Machine review of arXiv:2502.08200}
}

read the original abstract

Precise classification of megakaryocytes is crucial for diagnosing myelodysplastic syndromes. Although self-supervised learning has shown promise in medical image analysis, its application to classifying megakaryocytes in stained slides faces three main challenges: (1) pervasive background noise that obscures cellular details, (2) a long-tailed distribution that limits data for rare subtypes, and (3) complex morphological variations leading to high intra-class variability. To address these issues, we propose the ActiveSSF framework, which integrates active learning with self-supervised pretraining. Specifically, our approach employs Gaussian filtering combined with K-means clustering and HSV analysis (augmented by clinical prior knowledge) for accurate region-of-interest extraction; an adaptive sample selection mechanism that dynamically adjusts similarity thresholds to mitigate class imbalance; and prototype clustering on labeled samples to overcome morphological complexity. Experimental results on clinical megakaryocyte datasets demonstrate that ActiveSSF not only achieves state-of-the-art performance but also significantly improves recognition accuracy for rare subtypes. Moreover, the integration of these advanced techniques further underscores the practical potential of ActiveSSF in clinical settings.

Figures

Figures reproduced from arXiv: 2502.08200 by the authors.

**Figure 2.** We present a two-stage active learning framework for guiding self-supervised pretraining. Stage 1 (Cell Region [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Comparison of classification performance across all [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 10 canonical work pages

[1]

Identity mappings in deep residual networks,

K. He, X. Zhang, S. Ren, and J. Sun, “Identity mappings in deep residual networks,” in Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11– 14, 2016, Proceedings, Part IV 14 . Springer, 2016, pp. 630–645

work page 2016
[2]

Some methods for classification and analysis of multivariate observations,

J. MacQueen, “Some methods for classification and analysis of multivariate observations,” in Proceedings of 5-th Berkeley Sym- posium on Mathematical Statistics and Probability/University of California Press, 1967

work page 1967
[3]

A neural-network-based approach to white blood cell classification,

M.-C. Su, C.-Y . Cheng, and P.-C. Wang, “A neural-network-based approach to white blood cell classification,” The scientific world journal, vol. 2014, no. 1, p. 796371, 2014

work page 2014
[4]

Computer aided system for red blood cell classification in blood smear image,

R. Tomari, W. N. W. Zakaria, M. M. A. Jamil, F. M. Nor, and N. F. N. Fuad, “Computer aided system for red blood cell classification in blood smear image,” Procedia Computer Science , vol. 42, pp. 206–213, 2014, medical and Rehabilitation Robotics and Instrumentation (MRRI2013). [Online]. Available: https: //www.sciencedirect.com/science/article/pii/S18770...

work page 2014
[5]

Fusing pre-trained convolutional neural net- works features for multi-differentiated subtypes of liver cancer on histopathological images,

X. Dong, M. Li, P. Zhou, X. Deng, S. Li, X. Zhao, Y . Wu, J. Qin, and W. Guo, “Fusing pre-trained convolutional neural net- works features for multi-differentiated subtypes of liver cancer on histopathological images,” BMC Medical Informatics and Decision Making, vol. 22, no. 1, p. 122, 2022

work page 2022
[6]

Ultrasound nodule segmentation using asymmetric learning with simple clinical annotation,

X. Zhao, Z. Li, X. Luo, P. Li, P. Huang, J. Zhu, Y . Liu, J. Zhu, M. Yang, S. Chang, and J. Dong, “Ultrasound nodule segmentation using asymmetric learning with simple clinical annotation,” IEEE Transactions on Circuits and Systems for Video Technology , pp. 1–1, 2024

work page 2024
[7]

Sam- driven weakly supervised nodule segmentation with uncertainty- aware cross teaching,

X. Zhao, P. Li, X. Luo, M. Yang, S. Chang, and Z. Li, “Sam- driven weakly supervised nodule segmentation with uncertainty- aware cross teaching,” in 2024 IEEE International Symposium on Biomedical Imaging (ISBI) , 2024, pp. 1–5

work page 2024
[8]

Hfgs: 4d gaussian splatting with emphasis on spatial and temporal high-frequency components for endoscopic scene reconstruction,

H. Zhao, X. Zhao, L. Zhu, W. Zheng, and Y . Xu, “Hfgs: 4d gaussian splatting with emphasis on spatial and temporal high-frequency components for endoscopic scene reconstruction,” arXiv preprint arXiv:2405.17872, 2024

arXiv 2024

Show all 15 references

[9]

D-lmbmap: a fully automated deep-learning pipeline for whole-brain profiling of neural circuitry,

Z. Li, Z. Shang, J. Liu, H. Zhen, E. Zhu, S. Zhong, R. N. Sturgess, Y . Zhou, X. Hu, X. Zhao et al. , “D-lmbmap: a fully automated deep-learning pipeline for whole-brain profiling of neural circuitry,” Nature Methods, vol. 20, no. 10, pp. 1593–1604, 2023

2023
[10]

An empirical study of training self- supervised vision transformers,

X. Chen, S. Xie, and K. He, “An empirical study of training self- supervised vision transformers,” in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 9640–9649

2021
[11]

Convmae: Masked convolution meets masked autoencoders,

P. Gao, T. Ma, H. Li, Z. Lin, J. Dai, and Y . Qiao, “Convmae: Masked convolution meets masked autoencoders,” arXiv preprint arXiv:2205.03892, 2022

2022 arXiv
[12]

Mixed autoencoder for self-supervised visual representation learning,

K. Chen, Z. Liu, L. Hong, H. Xu, Z. Li, and D.-Y . Yeung, “Mixed autoencoder for self-supervised visual representation learning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2023, pp. 22 742–22 751

2023
[13]

Masked autoencoders are scalable vision learners,

K. He, X. Chen, S. Xie, Y . Li, P. Doll ´ar, and R. Girshick, “Masked autoencoders are scalable vision learners,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 16 000–16 009

2022
[14]

Momentum contrast for unsupervised visual representation learning,

K. He, H. Fan, Y . Wu, S. Xie, and R. Girshick, “Momentum contrast for unsupervised visual representation learning,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 9729–9738

2020
[15]

An image is worth 16x16 words: Transformers for image recognition at scale,

D. Alexey, “An image is worth 16x16 words: Transformers for image recognition at scale,” arXiv preprint arXiv: 2010.11929 , 2020

2010 arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

Identity mappings in deep residual networks,

K. He, X. Zhang, S. Ren, and J. Sun, “Identity mappings in deep residual networks,” in Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11– 14, 2016, Proceedings, Part IV 14 . Springer, 2016, pp. 630–645

work page 2016

[2] [2]

Some methods for classification and analysis of multivariate observations,

J. MacQueen, “Some methods for classification and analysis of multivariate observations,” in Proceedings of 5-th Berkeley Sym- posium on Mathematical Statistics and Probability/University of California Press, 1967

work page 1967

[3] [3]

A neural-network-based approach to white blood cell classification,

M.-C. Su, C.-Y . Cheng, and P.-C. Wang, “A neural-network-based approach to white blood cell classification,” The scientific world journal, vol. 2014, no. 1, p. 796371, 2014

work page 2014

[4] [4]

Computer aided system for red blood cell classification in blood smear image,

R. Tomari, W. N. W. Zakaria, M. M. A. Jamil, F. M. Nor, and N. F. N. Fuad, “Computer aided system for red blood cell classification in blood smear image,” Procedia Computer Science , vol. 42, pp. 206–213, 2014, medical and Rehabilitation Robotics and Instrumentation (MRRI2013). [Online]. Available: https: //www.sciencedirect.com/science/article/pii/S18770...

work page 2014

[5] [5]

Fusing pre-trained convolutional neural net- works features for multi-differentiated subtypes of liver cancer on histopathological images,

X. Dong, M. Li, P. Zhou, X. Deng, S. Li, X. Zhao, Y . Wu, J. Qin, and W. Guo, “Fusing pre-trained convolutional neural net- works features for multi-differentiated subtypes of liver cancer on histopathological images,” BMC Medical Informatics and Decision Making, vol. 22, no. 1, p. 122, 2022

work page 2022

[6] [6]

Ultrasound nodule segmentation using asymmetric learning with simple clinical annotation,

X. Zhao, Z. Li, X. Luo, P. Li, P. Huang, J. Zhu, Y . Liu, J. Zhu, M. Yang, S. Chang, and J. Dong, “Ultrasound nodule segmentation using asymmetric learning with simple clinical annotation,” IEEE Transactions on Circuits and Systems for Video Technology , pp. 1–1, 2024

work page 2024

[7] [7]

Sam- driven weakly supervised nodule segmentation with uncertainty- aware cross teaching,

X. Zhao, P. Li, X. Luo, M. Yang, S. Chang, and Z. Li, “Sam- driven weakly supervised nodule segmentation with uncertainty- aware cross teaching,” in 2024 IEEE International Symposium on Biomedical Imaging (ISBI) , 2024, pp. 1–5

work page 2024

[8] [8]

Hfgs: 4d gaussian splatting with emphasis on spatial and temporal high-frequency components for endoscopic scene reconstruction,

H. Zhao, X. Zhao, L. Zhu, W. Zheng, and Y . Xu, “Hfgs: 4d gaussian splatting with emphasis on spatial and temporal high-frequency components for endoscopic scene reconstruction,” arXiv preprint arXiv:2405.17872, 2024

arXiv 2024