A Novel Downsampling Strategy Based on Information Complementarity for Medical Image Segmentation

Chang Li; Guoping Xu; Wenbo Yue

REVIEW 5 major objections 5 minor 29 references

A Novel Downsampling Strategy Based on Information Complementarity for Medical Image Segmentation

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that replacing standard downsampling with min-plus-max pooling improves medical image segmentation Dice by about 0.5% on average.

desk verdict A simple min-plus-max pooling idea that is drowned by unverifiable numbers—no error bars, contradictory splits, and small inconsistent gains. read the letter →

arxiv 2507.14790 v1 pith:VZM6UFA3 submitted 2025-07-20 cs.CV

Wenbo Yue , Chang Li , Guoping Xu This is my paper · ORCID

classification cs.CV

keywords DownsamplingSemanticsegmentationMinMaxPoolingInformationcomplementarityMedicalimageDicesimilaritycoefficientHybridACDCdataset

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

This paper claims that information loss in CNN downsampling can be reduced by replacing max pooling or strided convolution with a hybrid pooling step that adds the minimum and maximum values of each local window. The proposed HPD module applies this MinMaxPooling operation and then passes the result through 1x1 convolution, batch normalization, and ReLU. On the ACDC cardiac and Synapse abdominal datasets, swapping HPD into six segmentation architectures improves mean Dice by about 0.5% on average, with several per-organ gains larger. The claim matters because downsampling is a fixed and cheap component of most segmentation networks, so a drop-in replacement that preserves more boundary and texture information could improve medical image segmentation without redesigning the network.

What carries the argument

The load-bearing object is the MinMaxPooling transform of Eq. (3): for each k×k window i, F_i = M_1(i) + M_2(i), where M_1(i) is the minimum of the window and M_2(i) the maximum. This is a named operation in the paper; it replaces the pooled value in standard downsampling and is followed by a feature learning module of 1×1 convolution, batch normalization, and ReLU that adjusts channels and filters redundancy. The min-max sum is claimed to behave as a combination of low-pass and high-pass filters, preserving light-dark contrast that max pooling alone discards.

What would settle it

Retrain the paper's baseline architectures on ACDC and Synapse with HPD replaced by max-average pooling or by plain max pooling with the same extra 1×1 convolution and channel adjustment, keeping parameters and FLOPs matched. If the mean Dice advantage of HPD shrinks to within random variation across several seeds, the claim that min-max summation specifically preserves complementary boundary information is not supported.

Watch

Extended reading notes

Core claim

The central claim is that a downsampling layer whose output is the sum of the minimum and the maximum within each pooling window retains complementary information -- the minimum acts as a low-pass component capturing overall local intensity, the maximum as a high-pass component capturing salient detail -- and this retained contrast improves dense pixel prediction. The paper reports that on ACDC, U-Net mDSC rises from 0.8993 to 0.9032, LinkNet from 0.869 to 0.8767, and TransUNet from 0.8842 to 0.8920, while Deeplabv3+ drops from 0.8001 to 0.7901; on Synapse, U-Net rises from 0.7596 to 0.7638 and LinkNet from 0.7474 to 0.7578. The authors interpret the Deeplabv3+ drop as redundancy with a similar mechanism already present in that architecture.

Load-bearing premise

The load-bearing premise is that adding the minimum value in each pooling window to the maximum value preserves boundary and contrast information relevant to segmentation, and that the small observed Dice gains come from this fusion rather than from the extra parameters and FLOPs the HPD module introduces.

Editorial extensions

If this is right

HPD is designed as a direct replacement for existing downsampling layers, so U-Net, LinkNet, Fast-SCNN, Swin-Unet, and TransUNet can adopt it without changing loss, optimizer, or training schedule.
On the reported ACDC numbers, mean Dice improves for five of six architectures; the largest relative gains appear in boundary-sensitive structures such as myocardium and the right ventricle.
On Synapse, per-organ gains can be much larger than the average, e.g. U-Net pancreas Dice rises from 0.5267 to 0.5929.
Parameter and FLOP counts rise only modestly (e.g., U-Net from 14.79M to 18.70M parameters and 31.05G to 35.27G FLOPs), so the reported gains are bought with a small compute increase.
Deeplabv3+ degrades with HPD, which the paper attributes to redundancy with an existing complementation mechanism; a practical implication is that HPD should be applied selectively per architecture.

Reading between the lines

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

A natural extension the paper does not run is an ablation comparing MinMaxPooling against max-average pooling or stochastic pooling with equal parameter counts; that would isolate whether the min term specifically causes the gains.
The reported gains are averages over datasets, and the paper does not report variance or significance tests; whether the 0.5% average is stable across seeds is testable by re-running the same tables.
If the low-pass/high-pass interpretation is right, HPD should generalize to other dense prediction tasks such as depth estimation or panoptic segmentation, and to datasets with small structures; that is an implication, not a claim the paper makes.
The 'information complementarity' concept could be pushed further by learning per-window weights for the min and max branches instead of a fixed sum.

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 read of the HPD paper. The claim is that summing min-pooling and max-pooling before downsampling gives a ~0.5% mDSC gain across segmentation architectures. The idea is simple enough that it might work, but this paper doesn't show it. The experiments cover six architectures on ACDC and five on Synapse, and the authors include a parameter/FLOP table, which is a reasonable amount of work. Credit where due: the operation is cheap, easy to drop into any CNN, and the breadth of architectures is okay.

But the evidence is not credible. There are no error bars, no multi-seed runs, no significance tests. The deltas are small (0.2-0.9%) and some are negative (DeepLabv3+ on ACDC, Fast-SCNN on Synapse), which the text admits. The ACDC split is given three different ways: 1312/101/380 in Section III.A, 1312/380/210 in Table I, and 70/20/10 in Table II. That is not a typo-level inconsistency; it means the experiments cannot be reproduced from the paper. The theoretical story about min pooling as low-pass and max pooling as high-pass is asserted, not analyzed; in a post-ReLU feature map, the min is just a positive offset added to the max, and there is no evidence this preserves boundary information. The ablation on number of HPD modules is non-monotonic, and four HPD modules is actually worse than zero, which undercuts the 'more is better' reading. No code is released. The references do not include any prior work on mixed or hybrid pooling, so the novelty claim is unproven.

On the positive side, the paper is honest enough to report the negative results in the tables, and the framing about information complementarity is a reasonable motivation even if it is not quantified. My take: this is a small idea with sloppy execution. I would not send it to a referee as it stands; a desk reject with an invitation to resubmit a properly evaluated version (with code, seeds, variance, corrected split numbers, and comparisons to other pooling variants) would be appropriate.

Referee Report

5 major / 5 minor

Summary. This paper proposes Hybrid Pooling Downsampling (HPD), a drop-in replacement for max pooling in which each pooling window is aggregated by the sum of its minimum and maximum values, followed by a 1x1 convolution, batch normalization, and ReLU. The authors evaluate HPD on ACDC and Synapse across six and five segmentation architectures respectively and claim an average improvement of 0.5% in mean Dice similarity coefficient (mDSC). They also report parameter/FLOP counts and a small ablation on the number of HPD modules in UNet. The central claim is that HPD outperforms traditional downsampling due to information complementarity of the min and max components.

Significance. If the reported gains were robust, HPD would be a simple and attractive contribution: it is a fixed arithmetic operation requiring no learned parameters beyond the existing 1x1 conv, it can be inserted into any CNN, and the authors include a computational-cost comparison (Table III). The paper also provides visualizations suggesting finer boundary detail. However, the evidence presented does not currently establish the central claim. The effect sizes are small and noisy, the paper's own tables contain counterexamples to the uniform-improvement claim, the dataset split is inconsistent across the text and tables, and the theoretical explanation in Section II.A is an unsupported analogy rather than an analysis. The contribution is therefore not yet validated.

major comments (5)

[Section III.C/Table IV and Section III.D/Table VI] The central claim that HPD 'outperforms traditional methods' is contradicted by the paper's own results: DeepLabv3+_MinMax drops from 0.8001 to 0.7901 mDSC on ACDC and Fast-SCNN_MinMax drops from 0.6487 to 0.6465 mDSC on Synapse, with several per-organ scores also decreasing. The abstract's stated average improvement of 0.5% is not traceable to the tables because the aggregation method is unspecified and no significance test is reported. These counterexamples and the mismatch with the abstract are load-bearing for the central claim.
[Section III.A and Tables I/II] The ACDC data split is irreproducible. The text states 1312/101/380 train/validation/test images, Table I lists 1312/380/210, and Table II lists 70/20/10 without a unit; the stated total of 1792 images matches none of these partitions (1312+380+210=1902, 1312+101+380=1793). The Synapse split is also described inconsistently (18/6/6 cases in the text vs. 2211/764/804 slices in Table I). Without a consistent and clearly stated split, the reported mDSC values cannot be reproduced or compared with prior literature.
[Section II.A, Eq. (3)] The theoretical basis of HPD is the assertion that F = M_min + M_max preserves complementary low- and high-frequency information, with min pooling acting as a low-pass filter and max pooling as a high-pass filter. No derivation, feature-space measurement, or comparison with alternative pooling combinations (e.g., avg+max, min+avg, or L_p pooling) is provided. In a post-ReLU network the min component is a nonnegative per-window offset, and it is unclear why this offset should retain boundary information beyond max pooling. This unsupported analogy is load-bearing because it motivates the entire method.
[Section III.B/Table III] The claim that HPD does not significantly increase computational cost is contradicted by Table III for UNet, where parameters increase from 14.79M to 18.70M (+26%) and FLOPs from 31.05G to 35.27G. Moreover, HPD adds a 1x1 convolution, BN, and ReLU after the pooling operation, while the baselines use their original downsampling; therefore, any observed gains may be due to the additional capacity rather than the min+max aggregation. An ablation that appends the same 1x1 conv/BN/ReLU to max pooling is needed to isolate the effect.
[All experiments] No error bars, multiple seeds, or significance tests are reported. The observed deltas (e.g., UNet +0.39% on ACDC, TransUNet +0.17% on Synapse) are within the typical run-to-run variance of medical segmentation models, so the claimed 0.5% average improvement is not distinguishable from noise. Reporting the mean and standard deviation over at least three runs per configuration, or a paired test, is necessary to support the central claim.

minor comments (5)

[Throughout] The text contains multiple typos and inconsistencies: 'Hybid' in Section II, 'minmum' in Section II.A, and inconsistent spacing in 'minmaxpooling' throughout.
[Section III.B] The sentence 'The word “data” is plural, not singular.' is unrelated to the technical content and should be removed.
[References] Reference [25] is cited as 'the method of Chen et al.' but the listed reference is a remote-sensing change-detection paper with different authors; the correct citation for the slice-by-slice training protocol is missing.
[Table IV] The note says 'The best result is marked in bold', but no entries are typeset in bold.
[Eq. (4)] Equation (4) is garbled in the text; if it is intended to be DSC = 2TP/(2TP+FP+FN), the typesetting should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: HPD is a fixed arithmetic operation and the reported Dice gains are empirical measurements, not outputs of a derivation assuming the conclusion.

full rationale

The paper's central claim is that replacing standard downsampling with the fixed operation F = M_min + M_max (Eq. 3) improves mDSC by about 0.5% on average across several architectures. The definition of HPD does not encode or assume the experimental outcome: MinMaxPooling is a parameter-free arithmetic combination of the min and max over each local window, and the reported Dice scores in Tables IV and VI are measurements from trained models, not quantities recovered from the method's definition. There is no fitted parameter that is later renamed a prediction, no equation that reduces to its own input, and no uniqueness theorem or prior result invoked to force the choice. The low-pass/high-pass analogy for min and max pooling is a post hoc narrative, not a derivation step from which the segmentation gains follow deductively. The paper does cite prior work by its own authors (e.g., reference [8] LCU-Net and reference [17] DiSegNet), but these appear only as related-work examples and are not load-bearing for the HPD claim. The serious weaknesses in the paper are experimental, not circular: the abstract says ACDC has 1312/101/380 train/validation/test images while Table I reports 1312/380/210 and Table II reports 70/20/10, no standard deviations or significance tests are provided, and several per-model deltas are negative or within typical run-to-run variation. These issues affect verifiability and correctness, but they do not make the derivation circular. The claimed improvement is an empirical assertion measured against a baseline, and the method's definition does not presuppose that assertion. Therefore no circularity is present.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new physical or mathematical entities. HPD is a composition of existing operations, and the only free choices are the unspecified pooling window size and the number of HPD modules, neither of which is fitted to data in a reproducible way.

free parameters (2)

Pooling window size k
Eqs. (1)-(3) define pooling over windows of size k, but k is never assigned a value in the paper. The text says the output is half resolution, which implies k=2, but this is not stated explicitly.
Number of HPD modules per architecture
The ablation in Table V varies the count from 0 to 4 and reports different Dice values. The final choice of where and how many HPD modules replace downsampling is not governed by a stated selection rule.

assumptions (3)

standard math Standard CNN operations (convolution, batch normalization, ReLU, pooling) and backpropagation work as usual
The method is built on these operations without modification.
domain assumption Minimum pooling acts as a low-pass filter and maximum pooling as a high-pass filter, and their sum preserves light-dark contrast and detail
Assumed in Section II.A around Eq. (3); no experimental or theoretical evidence is provided beyond visual examples.
domain assumption The reported Dice scores come from a reliable training setup, a single run is representative, and the dataset splits are correct
The paper relies on single-run comparisons, and the dataset split numbers are internally inconsistent between the text and Table I.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Novel Downsampling Strategy Based on Information Complementarity for Medical Image Segmentation." pith.science (2026). https://pith.science/paper/VZM6UFA3

@misc{pith2026250714790,
  author       = {Pith},
  title        = {Pith review of: A Novel Downsampling Strategy Based on Information Complementarity for Medical Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VZM6UFA3}},
  note         = {Machine review of arXiv:2507.14790}
}

read the original abstract

In convolutional neural networks (CNNs), downsampling operations are crucial to model performance. Although traditional downsampling methods (such as maximum pooling and cross-row convolution) perform well in feature aggregation, receptive field expansion, and computational reduction, they may lead to the loss of key spatial information in semantic segmentation tasks, thereby affecting the pixel-by-pixel prediction accuracy.To this end, this study proposes a downsampling method based on information complementarity - Hybrid Pooling Downsampling (HPD). The core is to replace the traditional method with MinMaxPooling, and effectively retain the light and dark contrast and detail features of the image by extracting the maximum value information of the local area.Experiment on various CNN architectures on the ACDC and Synapse datasets show that HPD outperforms traditional methods in segmentation performance, and increases the DSC coefficient by 0.5% on average. The results show that the HPD module provides an efficient solution for semantic segmentation tasks.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 25 canonical work pages

[1]

Deformation models for image recognition[J]

Keysers D, Deselaers T, Gollan C, et al. Deformation models for image recognition[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2007, 29(8): 1422-1435

work page 2007
[2]

Object detection with deep learning: A review[J]

Zhao Z Q, Zheng P, Xu S, et al. Object detection with deep learning: A review[J]. IEEE transactions on neural networks and learning systems, 2019, 30(11): 3212-3232

work page 2019
[3]

Review the state-of-the-art technologies of semantic segmentation based on deep learning[J]

Mo Y, Wu Y, Yang X, et al. Review the state-of-the-art technologies of semantic segmentation based on deep learning[J]. Neurocomputing, 2022, 493: 626-646

work page 2022
[4]

Imagenet classification with deep convolutional neural networks[J]

Krizhevsky A, Sutskever I, Hinton G E. Imagenet classification with deep convolutional neural networks[J]. Advances in neural information processing systems, 2012, 25

work page 2012
[5]

Going deeper with convolutions[C]//Proceedings of the IEEE conference on computer vision and pattern recognition

Szegedy C, Liu W, Jia Y, et al. Going deeper with convolutions[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2015: 1-9

work page 2015
[6]

Deep residual learning for image recognition[C]//Proceedings of the IEEE conference on computer vision and pattern recognition

He K, Zhang X, Ren S, et al. Deep residual learning for image recognition[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2016: 770-778

2016
[7]

Ronneberger O, Fischer P, Brox T. U-net: Convolutional networks for biomedical image segmentation[C]//Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18. Springer International Publishing, 2015: 234-241.\

work page 2015
[8]

LCU-Net: A novel low-cost U-Net for environmental microorganism image segmentation[J]

Zhang J, Li C, Kosov S, et al. LCU-Net: A novel low-cost U-Net for environmental microorganism image segmentation[J]. Pattern Recognition, 2021, 115: 107885

work page 2021

Show all 29 references

[9]

Contextual ensemble network for semantic segmentation[J]

Zhou Q, Wu X, Zhang S, et al. Contextual ensemble network for semantic segmentation[J]. Pattern Recognition, 2022, 122: 108290

2022
[10]

Linknet: Exploiting encoder representations for efficient semantic segmentation[C]//2017 IEEE visual communications and image processing (VCIP)

Chaurasia A, Culurciello E. Linknet: Exploiting encoder representations for efficient semantic segmentation[C]//2017 IEEE visual communications and image processing (VCIP). IEEE, 2017: 1- 4

2017
[11]

Refinenet: Multi-path refinement networks for high-resolution semantic segmentation[C]//Proceedings of the IEEE conference on computer vision and pattern recognition

Lin G, Milan A, Shen C, et al. Refinenet: Multi-path refinement networks for high-resolution semantic segmentation[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 1925-1934

2017
[12]

Encoder-decoder with atrous separable convolution for semantic image segmentation[C]//Proceedings of the European conference on computer vision (ECCV)

Chen L C, Zhu Y, Papandreou G, et al. Encoder-decoder with atrous separable convolution for semantic image segmentation[C]//Proceedings of the European conference on computer vision (ECCV). 2018: 801-818

2018
[13]

Pyramid scene parsing network[C]//Proceedings of the IEEE conference on computer vision and pattern recognition

Zhao H, Shi J, Qi X, et al. Pyramid scene parsing network[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 2881-2890

2017
[14]

Progressive global perception and local polishing network for lung infection segmentation of COVID- 19 CT images[J]

Mu N, Wang H, Zhang Y, et al. Progressive global perception and local polishing network for lung infection segmentation of COVID- 19 CT images[J]. Pattern Recognition, 2021, 120: 108168

2021
[15]

Deep high-resolution representation learning for visual recognition[J]

Wang J, Sun K, Cheng T, et al. Deep high-resolution representation learning for visual recognition[J]. IEEE transactions on pattern analysis and machine intelligence, 2020, 43(10): 3349-3364

2020
[16]

Icnet for real-time semantic segmentation on high-resolution images[C]//Proceedings of the European conference on computer vision (ECCV)

Zhao H, Qi X, Shen X, et al. Icnet for real-time semantic segmentation on high-resolution images[C]//Proceedings of the European conference on computer vision (ECCV). 2018: 405-420

2018
[17]

DiSegNet: A deep dilated convolutional encoder-decoder architecture for lymph node segmentation on PET/CT images[J]

Xu G, Cao H, Udupa J K, et al. DiSegNet: A deep dilated convolutional encoder-decoder architecture for lymph node segmentation on PET/CT images[J]. Computerized Medical Imaging and Graphics, 2021, 88: 101851

2021
[18]

Multi-modal unsupervised domain adaptation for semantic image segmentation[J]

Hu S, Bonardi F, Bouchafa S, et al. Multi-modal unsupervised domain adaptation for semantic image segmentation[J]. Pattern Recognition, 2023, 137: 109299

2023
[19]

CANet: Co-attention network for RGB-D semantic segmentation[J]

Zhou H, Qi L, Huang H, et al. CANet: Co-attention network for RGB-D semantic segmentation[J]. Pattern Recognition, 2022, 124: 108468

2022
[20]

Rgbd-net: Predicting color and depth images for novel views synthesis[C]//2021 International Conference on 3D Vision (3DV)

Nguyen P, Karnewar A, Huynh L, et al. Rgbd-net: Predicting color and depth images for novel views synthesis[C]//2021 International Conference on 3D Vision (3DV). IEEE, 2021: 1095-1105

2021
[21]

Semantic segmentation with boundary neural fields[C]//Proceedings of the IEEE conference on computer vision and pattern recognition

Bertasius G, Shi J, Torresani L. Semantic segmentation with boundary neural fields[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2016: 3602-3610

2016
[22]

Edge detection techniques-an overview[J]

Ziou D, Tabbone S. Edge detection techniques-an overview[J]. Распознавание образов и анализ изображен/Pattern Recognition and Image Analysis: Advances in Mathematical Theory and Applications, 1998, 8(4): 537-559

1998
[23]

A color image segmentation method as used in the study of ancient monument decay[J]

Cossu R, Chiappini L. A color image segmentation method as used in the study of ancient monument decay[J]. Journal of Cultural Heritage, 2004, 5(4): 385-391

2004
[24]

Terahertz plasmonic high pass filter[J]

Wu D, Fang N, Sun C, et al. Terahertz plasmonic high pass filter[J]. Applied Physics Letters, 2003, 83(1): 201-203

2003
[25]

X. Li, M. He, H. Li, H. Shen, A combined loss-based multiscale fully convolutional network for high-resolution remote sensing image change detection, IEEE Geosci.Remote Sens. Lett. 19 (2021) 1–5

2021
[26]

Fast-scnn: Fast semantic segmentation network[J]

Poudel R P K, Liwicki S, Cipolla R. Fast-scnn: Fast semantic segmentation network[J]. arXiv preprint arXiv:1902.04502, 2019

1902 arXiv
[27]

Rethinking atrous convolution for semantic image segmentation[J]

Chen L C. Rethinking atrous convolution for semantic image segmentation[J]. arXiv preprint arXiv:1706.05587, 2017

2017 arXiv
[28]

Swin-unet: Unet-like pure transformer for medical image segmentation[C]//European conference on computer vision

Cao H, Wang Y, Chen J, et al. Swin-unet: Unet-like pure transformer for medical image segmentation[C]//European conference on computer vision. Cham: Springer Nature Switzerland, 2022: 205-218

2022
[29]

Transunet: Transformers make strong encoders for medical image segmentation[J]

Chen J, Lu Y, Yu Q, et al. Transunet: Transformers make strong encoders for medical image segmentation[J]. arXiv preprint arXiv:2102.04306, 2021

2021 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Deformation models for image recognition[J]

Keysers D, Deselaers T, Gollan C, et al. Deformation models for image recognition[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2007, 29(8): 1422-1435

work page 2007

[2] [2]

Object detection with deep learning: A review[J]

Zhao Z Q, Zheng P, Xu S, et al. Object detection with deep learning: A review[J]. IEEE transactions on neural networks and learning systems, 2019, 30(11): 3212-3232

work page 2019

[3] [3]

Review the state-of-the-art technologies of semantic segmentation based on deep learning[J]

Mo Y, Wu Y, Yang X, et al. Review the state-of-the-art technologies of semantic segmentation based on deep learning[J]. Neurocomputing, 2022, 493: 626-646

work page 2022

[4] [4]

Imagenet classification with deep convolutional neural networks[J]

Krizhevsky A, Sutskever I, Hinton G E. Imagenet classification with deep convolutional neural networks[J]. Advances in neural information processing systems, 2012, 25

work page 2012

[5] [5]

Going deeper with convolutions[C]//Proceedings of the IEEE conference on computer vision and pattern recognition

Szegedy C, Liu W, Jia Y, et al. Going deeper with convolutions[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2015: 1-9

work page 2015

[6] [6]

Deep residual learning for image recognition[C]//Proceedings of the IEEE conference on computer vision and pattern recognition

He K, Zhang X, Ren S, et al. Deep residual learning for image recognition[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2016: 770-778

2016

[7] [7]

Ronneberger O, Fischer P, Brox T. U-net: Convolutional networks for biomedical image segmentation[C]//Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18. Springer International Publishing, 2015: 234-241.\

work page 2015

[8] [8]

LCU-Net: A novel low-cost U-Net for environmental microorganism image segmentation[J]

Zhang J, Li C, Kosov S, et al. LCU-Net: A novel low-cost U-Net for environmental microorganism image segmentation[J]. Pattern Recognition, 2021, 115: 107885

work page 2021