Deep-Learning-Assisted Highly-Accurate COVID-19 Diagnosis on Lung Computed Tomography Images

Juhyun Bae; Ka Ho Chow; Shenyang Chen; Shreyash Gupta; Yinuo Wang

REVIEW 4 major objections 6 minor 23 references

Deep-Learning-Assisted Highly-Accurate COVID-19 Diagnosis on Lung Computed Tomography Images

T0 review · 4 major / 6 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 A combined GAN-cleaning and loss-adjustment pipeline reports over 0.983 MCC for four-class COVID-19 severity and 100% patient-level accuracy via voting.

desk verdict The pipeline is a reasonable engineering effort, but patient leakage and incorrect metrics invalidate the headline accuracy claims. read the letter →

arxiv 2507.04252 v1 pith:55TA3NAM submitted 2025-07-06 eess.IV cs.AIcs.CV

Yinuo Wang , Juhyun Bae , Ka Ho Chow , Shenyang Chen , Shreyash Gupta This is my paper · ORCID

classification eess.IVcs.AIcs.CV

keywords COVID-19CTimagingMosMedDataResNet18generativeadversarialnetworksLDAMlossclass-balancedimbalance

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 an end-to-end deep-learning pipeline can classify COVID-19 severity from lung CT slices with near-perfect accuracy, even when the training data are heavily imbalanced. It builds the pipeline from a ResNet18 backbone, a GAN-based quality-control stage that standardizes slices, imputes missing image regions, and flags anomalies, and two class-sensitive loss functions (LDAM and class-balanced loss) applied together. On the MosMedData collection, the paper reports a Matthews correlation coefficient above 0.983 on the benchmark test set and, after majority voting over ten slices per patient, 100% accuracy at the patient level. If these numbers hold with proper evaluation, such a system could support CT-based COVID-19 screening and severity grading where RT-PCR is stressed.

What carries the argument

The system's core is a ResNet18 feature extractor preceded by a data-quality-control stage and followed by a voting stage. The quality-control stage selects lung-region layers, interpolates every patient to ten slices, applies min-max normalization to Hounsfield-unit values, repairs abnormally dark layers with a sliding window over adjacent slices, imputes missing regions with MisGAN, and uses AnoGAN to highlight anomalous areas. The loss stage combines label-distribution-aware margin (LDAM), which adds a per-class margin inversely related to the class count, with class-balanced reweighting based on the effective number of samples; the paper's grid search identifies the best margin coefficient and expanding factor. Patient-level diagnosis is then produced by majority voting over the ten slice predictions.

What would settle it

Re-run training and evaluation with a strict patient-level split of MosMedData, so no patient's slices appear in both training and test, and recompute the four-class MCC and the patient-level voting accuracy; if either falls below 0.983 or 100%, the central claim is contradicted.

Watch

Extended reading notes

Core claim

The central claim is that a conventional ResNet18 classifier, when fed CT slices cleaned by a GAN-based standardization and imputation pipeline and trained with LDAM plus class-balanced reweighting, outperforms both the plain baseline and the alternate imbalance strategies (focal loss, LDAM alone, class-balanced alone) on four-class COVID-19 severity classification. The authors report more than 0.983 MCC on their benchmark image-level test set and an F1 score of 99.12%, which they state beats earlier same-dataset results around 97-98% F1. They also claim that after voting over the ten preprocessed slices of a patient, the patient-level diagnosis accuracy reaches 100%, with the class receiving the most slice votes becoming the diagnosis. The paper frames the result as evidence that data quality control and loss-level imbalance handling matter more than backbone choice for imbalanced medical imaging.

Load-bearing premise

The evaluation treats individual CT slices from the same patient as independent, with random splits and fourfold cross-validation at the image level, so slices from one patient can appear in both training and test sets; if patients were held out instead, the reported 0.983 MCC and 100% patient-level accuracy may not hold.

Editorial extensions

If this is right

A four-class COVID-19 severity classifier on MosMedData can be built with a lightweight ResNet18 backbone plus preprocessing, reaching MCC above 0.983 on the paper's test set.
The combination of LDAM margin and class-balanced reweighting is the best-performing imbalance strategy among those compared, supporting the paper's claim that these two loss corrections are complementary.
GAN-based imputation and anomaly detection improve classification over the raw standardized slices, implying that corrupted or missing CT regions need not be discarded.
Majority voting over ten standardized slices yields the reported 100% patient-level diagnostic accuracy, so the system could in principle be deployed as a CT triage aid.

Reading between the lines

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

If the image-level random split and fourfold cross-validation place slices from the same patient in both training and test, the reported MCC and 100% patient-level accuracy are likely inflated; a strict patient-level split would be the decisive check.
The voting mechanism averages over ten spatially adjacent slices from one scan, so those votes are highly correlated; 100% accuracy on patients may reflect redundancy across slices rather than independent diagnostic evidence.
The same pipeline could be tested on other imbalanced chest CT datasets, such as pneumonia or lung cancer severity, to see whether the GAN-based quality-control and LDAM-plus-CB losses transfer.
A more informative report would give the distribution of vote margins for each patient, since perfect voting accuracy can hide near-ties between classes.

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

Short take: this paper's central claim—>0.983 MCC and 100% patient-level accuracy—is not supported by the evidence as reported. The evaluation leaks patients between train and test, and the metric formulas are wrong. That said, there is a real engineering effort here: the pipeline (standardization, MisGAN imputation, sliding-window artifact correction, and LDAM/CB/focal losses) is a sensible combination of known techniques, and the authors did compare against simpler baselines. The writing is clear and the references are mostly appropriate.

The main problem is the evaluation design. As you noted, Section VI.D says the model is "tested on the random split image-level test set," and the preprocessing outputs 10 slices per patient. A random image-level split does not keep patients together, so adjacent slices from the same patient end up in both training and testing. That inflates the MCC and makes the 100% patient-level voting accuracy trivial. This is a load-bearing flaw, not a tweak.

Second, the metrics are defined incorrectly. Eq. (18) is the binary MCC, not the multi-class version. Eq. (20) is not Cohen's kappa. Eq. (21) defines balanced accuracy as (TP+TN)/2, which is not sensitivity/specificity averaging. These aren't just typos; they affect the reported numbers.

Third, the headline numbers appear only in figures, not in any numeric table in the text. The abstract says "more than 0.983 MCC," but I couldn't find that value anywhere in the body. That's a transparency problem.

The paper also drops the CT4 class (only 2 patients) without discussion of how that affects the clinical claim, and the CT3 anomaly-score threshold (Table I) is selected post hoc on the same data. No code or data is released, despite the contact email saying "contact for datasets and source codes."

So the central argument doesn't hold. The pipeline might be useful to someone building a similar system, but the evaluation needs a complete rework with patient-level splitting and corrected metrics.

Who is this for? A reader interested in a concrete example of how not to evaluate a patient-level diagnostic model. It could be a good teaching case. But as a research contribution, it doesn't deserve a serious referee in its current form. I'd recommend rejecting it and suggesting the authors redo the evaluation properly, report numeric results by class, and release code.

Referee Report

4 major / 6 minor

Summary. The paper describes a deep-learning pipeline for four-class COVID-19 severity classification on lung CT slices from MosMedData. The pipeline combines a ResNet18 backbone with data quality control steps (layer selection and interpolation, min-max normalization, MisGAN-based missing-data imputation, and AnoGAN-based anomaly detection) and with long-tail losses (LDAM, focal, and class-balanced loss). The authors report that the model achieves more than 0.983 Matthews correlation coefficient (MCC) on the 'benchmark test dataset' and that a voting mechanism yields 100% patient-level diagnostic accuracy, but no numeric evaluation table is provided in the text.

Significance. If the reported near-perfect accuracy on a public dataset were supported by a sound, patient-disjoint evaluation, the work would be a useful contribution to automated COVID-19 severity grading and to the application of class-imbalance losses in medical imaging. The paper also demonstrates a comparison of three backbones and several loss-function combinations, which is of practical interest. However, the reported headline numbers are undermined by the evaluation design: the test protocol is an image-level random split that does not enforce patient independence, the central metrics are defined with binary formulas that do not apply to four-class classification, and the 'benchmark test dataset' is never defined as a separate patient-disjoint set. These issues affect the validity of every central quantitative claim.

major comments (4)

[§VI.B, §VI.D, Abstract] The evaluation does not enforce patient independence. Section IV.A states that the preprocessing pipeline outputs 10 CT slices for each patient, and Table II's training counts (2540, 6840, 1250, 450) are exactly ten times the dataset's patient counts (254, 684, 125, 45), confirming slice-level duplication. Section VI.B describes only fourfold cross-validation on training data, and Section VI.D says the model is 'test[ed] on the random split image-level test set.' A random image-level split places slices of the same patient in both training and test partitions; because adjacent slices from one volume are highly correlated and share the patient-level label, the reported MCC and the 100% patient-level accuracy are inflated and do not establish generalization to new patients. The abstract's 'benchmark test dataset' is never defined as a patient-disjoint held-out set, so the central claim is unsupported.
[§VI.B, Eqs. (18), (20), (21)] The metric definitions are incorrect for the four-class problem. Equation (18) defines MCC using a single TP, TN, FP, and FN, which is a binary-classification formula; it is not a valid multiclass MCC. Equation (21) defines balanced accuracy as (TP+TN)/2, again a binary formula, whereas the standard multiclass balanced accuracy is the average per-class recall. Equation (20) is labeled p0 (overall accuracy) but the formula given, sum(A_i * B_i)/n^2, is the expected agreement under independence (p_e in Cohen's kappa), not the observed agreement; the text itself states p0 should be the overall classification accuracy. Since the reported MCC, balanced accuracy, and kappa are central quantitative results, these formula errors invalidate the numerical claims as reported.
[§VI.D] The claim that 'the patient-level diagnostic result accuracy can reach 100%' is presented without any numerical support, without a description of the evaluation set, and without accounting for the image-level random split. Because each patient contributes 10 highly correlated slices and many of those patients' slices may appear in training under the random split, a majority-vote mechanism can produce artificially perfect agreement on patients already seen during training. No evidence is provided that this accuracy would transfer to unseen patients, so the claim is not established.
[§IV.C, Table I] The anomaly-detection component selects the CT3 threshold from the same data used to report detection performance. The text states that 'CT3 can be detected by threshold of Anomaly Score' based on the anomaly scores in Table I, but these scores are computed on the same test set used for evaluation, with no validation split or sensitivity analysis. This post-hoc threshold selection makes the anomaly-detection result circular and its contribution to the main pipeline unquantified.

minor comments (6)

[§VI.C, Figs. 12 and 13] The results in Figures 12 and 13 are reported only as bar charts; no numeric tables or confidence intervals are given in the text, which prevents verification of the abstract's 0.983 MCC claim and of the comparisons among models.
[§IV.A] The statement that 'the first and last 25% of layers' contain no diagnostically relevant tissue is presented as an assumption without supporting evidence or a reference; this should be justified or tested.
[Eq. (18)] The acronym is misspelled as 'MMC' in the equation; it should read 'MCC.'
[§IV.A] The term 'Hounsfiled Unit' should be 'Hounsfield Unit.'
[References] References [23]–[25] are duplicates of [5]–[7]; the duplicated entries should be removed or replaced with distinct citations.
[Throughout] The paper contains several typographical and grammatical errors (e.g., 'effect' for 'affect,' 'Alveolous,' 'distriminator,' 'groud true') that should be corrected in a revision.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper reports empirical measurements and contains no derived prediction that reduces to its own inputs.

full rationale

The paper is an empirical deep-learning study. Its headline claims (MCC above 0.983, patient-level accuracy of 100%) are measured outcomes of training and evaluation, not quantities derived from the model equations or from fitted constants renamed as predictions. The loss functions (LDAM, focal, CB), MisGAN-based imputation, and AnoGAN anomaly scoring are adopted from external references, and there is no load-bearing self-citation chain. The only potentially circular-looking element is Section IV.C, where an anomaly-score threshold is described after inspecting Table I; this threshold is not used to compute the reported MCC and is not presented as a predictive result, so it does not amount to a fitted input being called a prediction. The more serious issue is in Section VI.D, where the random image-level test split does not guarantee patient independence, so slices from the same patient may appear in both training and test sets; that is an evaluation-validity and generalization concern, not a circularity in the paper's derivation chain. No load-bearing step reduces, by construction, to the paper's own inputs.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

No invented physical entities; the ledger captures tuned hyperparameters and ad hoc preprocessing choices that the central accuracy claim depends on.

free parameters (7)

LDAM margin constant m = 0.3
Tuned via grid search on the same data; directly controls the margin in LDAM loss used in the final model.
LDAM expanding factor s = 50
Tuned via grid search; scales the logits before softmax.
Focal loss gamma = 0.02
Tuned via grid search, listed in system parameters; affects the weighting of easy versus hard samples.
Class-balanced loss beta = 0.999
Tuned via grid search; sets the effective number of samples in class-balanced reweighting.
Layer trim fraction = 25% (first and last layers removed)
Hand-chosen heuristic to isolate lung region; not validated against radiologist annotation.
Number of interpolated slices per patient = 10
Hand-chosen to unify layer counts; affects what the image-level classifier sees and the voting mechanism.
CT3 anomaly-score threshold = approximately 0.0355 (near mean CT3 score)
Chosen post hoc from Table I to separate CT3 from CT0-CT2; not evaluated on a held-out set.

assumptions (5)

domain assumption Every CT slice within a patient shares the patient-level label.
Stated in Section II; justifies slice-level training with patient-level labels, but ignores within-patient heterogeneity that the paper later acknowledges.
domain assumption Random image-level splits produce independent training and test points.
Implied by the evaluation design in Sections VI.B and VI.D; false because slices from one patient appear in both train and test.
ad hoc to paper The first and last 25% of CT layers contain no diagnostically relevant lung tissue.
Stated in Section IV.A; no evidence provided that this range selector preserves all lung-area slices.
domain assumption GAN-based imputation and interpolation do not introduce artifacts that change the classification target.
Invoked in Section IV.B; the paper only shows visual examples, no quantitative validation that synthetic imputation preserves diagnostic information.
standard math ResNet18 pre-training and the standard training recipe yield a well-calibrated classifier for this domain.
Used throughout; not proven for this dataset.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Deep-Learning-Assisted Highly-Accurate COVID-19 Diagnosis on Lung Computed Tomography Images." pith.science (2026). https://pith.science/paper/55TA3NAM

@misc{pith2026250704252,
  author       = {Pith},
  title        = {Pith review of: Deep-Learning-Assisted Highly-Accurate COVID-19 Diagnosis on Lung Computed Tomography Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/55TA3NAM}},
  note         = {Machine review of arXiv:2507.04252}
}

read the original abstract

COVID-19 is a severe and acute viral disease that can cause symptoms consistent with pneumonia in which inflammation is caused in the alveolous regions of the lungs leading to a build-up of fluid and breathing difficulties. Thus, the diagnosis of COVID using CT scans has been effective in assisting with RT-PCR diagnosis and severity classifications. In this paper, we proposed a new data quality control pipeline to refine the quality of CT images based on GAN and sliding windows. Also, we use class-sensitive cost functions including Label Distribution Aware Loss(LDAM Loss) and Class-balanced(CB) Loss to solve the long-tail problem existing in datasets. Our model reaches more than 0.983 MCC in the benchmark test dataset.

Figures

Figures reproduced from arXiv: 2507.04252 by the authors.

**Figure 1.** CT scans of MosMedData The dataset collecting 1110 patients’ CT was categorized into 5 patient-level labels based on severity of CT0 (Inconsistent with pneumonia, 254 samples, 22.8%), CT1 (Alveoli impacted in a scale of 25% or below, 684 samples, 61.6%), CT2 (25-50% alveoli impacted, 125 samples, 11.3%), CT3 (50-75% alveoli impacted, 45 samples, 4.1%) and CT4 (75% and above alveoli impacted, 2 samples, 0.2%). And e… view at source ↗

**Figure 2.** Data distribution of MosMedData It’s clear that there exists imbalance in the dataset. CT1 takes the majority among all the categories (about 61.6 %) while CT4 only takes a very small part in the dataset (about 0.2 %). In this paper, considering the difficulties to conduct multi-label classification under the conditions of imbalance, we ignore the CT4 category and focus on building up an endto-end machine learning … view at source ↗

**Figure 4.** Network Architecture of ResNet18 Similar to ResNet, GoogLeNet is a convolutional neural network which contains 22 layers [3]. And [3] proposes a new deep convolutional neural network architecture which is named as "Inception". The GoogLeNet can keep computational budget constant while reaching good classification performance. And another strong candidate, Vgg16, is also a deep convolutional neural network well-know… view at source ↗

Figures from the paper (10 more)

**Figure 5.** Figure 5: CT Images Standardization introduced by missing data, we take an imputation approach with MisGAN [8] to enhance the quality of CTs before sending to the classifier. MisGAN trains three GANs: (i) synthesizing CT images, (ii) synthesizing binary masks indicating the loca…

**Figure 6.** Figure 6: Missing Data Imputation C. Anomaly Detection The detection and quantification of disease markers in imaging data is important during diagnosis, and monitoring of disease progression, or treatment response. However, data may get corrupted during the data collection and …

**Figure 7.** Figure 7: Generated Fake Images and Anomaly Detection. from the left original image, generated image and anomaly detection for each CT1, CT2 and CT3. V. MEASUREMENTS TO SOLVE DATA IMBALANCE By analyzing the data distribution, it’s clear that the CT scan resources of different ca…

**Figure 8.** Figure 8: Heat map of parameter choices for ResNet backbone [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]

**Figure 9.** Figure 9: Heat map of parameter choices for GoogLeNet [PITH_FULL_IMAGE:figures/full_fig_p006_9.png]

**Figure 10.** Figure 10: Heat map of parameter choices for Vgg16 backbone [PITH_FULL_IMAGE:figures/full_fig_p006_10.png]

**Figure 11.** Figure 11: Workflow diagram The evaluation metrics are as following: Matthews correlation coefficient (MCC), F1 score, balanced accuracy, precision, sensitivity and Kappa. The focus is to more accurately describe the performance over the imbalanced datesets to avoid classifica…

**Figure 12.** Figure 12: Comparison of different models performance [PITH_FULL_IMAGE:figures/full_fig_p007_12.png]

**Figure 14.** Figure 14: Classification Result and Diagnostic Result [PITH_FULL_IMAGE:figures/full_fig_p008_14.png]

**Figure 13.** Figure 13: Comparison of ResNet performance with combina [PITH_FULL_IMAGE:figures/full_fig_p008_13.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 17 canonical work pages

[1]

Mosmeddata: Chest ct scans with covid-19 related findings dataset[J]

Morozov S P, Andreychenko A E, Pavlov N A, et al. Mosmeddata: Chest ct scans with covid-19 related findings dataset[J]. arXiv preprint arXiv:2005.06465, 2020

arXiv 2005
[2]

Deep residual learning[J]

He K, Zhang X, Ren S, et al. Deep residual learning[J]. Image Recognition, 2015

work page 2015
[3]

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

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

work page 2015
[4]

Very deep convolutional networks for large-scale image recognition[J]

Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition[J]. arXiv preprint arXiv:1409.1556, 2014

arXiv 2014
[8]

MisGAN: Learn- ing from Incomplete Data with Generative Adversarial Networks

Steven Cheng-Xian Li, Bo Jiang, Benjamin Marlin. "MisGAN: Learn- ing from Incomplete Data with Generative Adversarial Networks." International Conference on Learning Representations 2019

work page 2019
[9]

Improved training of wasserstein gans

Gulrajani, Ishaan, et al. "Improved training of wasserstein gans." Advances in neural information processing systems 30 (2017)

2017
[10]

Generative adversarial nets

Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y . "Generative adversarial nets." Advances in Neural Information Processing Systems. (2014)

work page 2014
[11]

Deep generative image models using a laplacian pyramid of adversarial networks

Denton, E.L., Chintala, S., Fergus, R., et al. "Deep generative image models using a laplacian pyramid of adversarial networks." Advances in neural information processing systems. (2015)

work page 2015

Show all 23 references

[12]

Unsupervised representation learning with deep convolutional generative adversarial networks

Radford, A., Metz, L., Chintala, S. "Unsupervised representation learning with deep convolutional generative adversarial networks." arXiv:1511.06434 (2015)

2015 arXiv
[13]

Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery

Schlegl et al. "Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery." IPMI 2017

2017
[14]

WHO COVID-19 Case defini- tion

World Health Organization (WHO). WHO COVID-19 Case defini- tion. WHO/2019-nCoV/Surveillance-Case-Definition/2020.1 2020 (ac- cessed 17 October 2020)

2019
[15]

Correlation of Chest CT and RT-PCR Testing for Coronavirus Disease 2019 (COVID-19) in China: A Report of 1014 Cases

Ai T, Yang Z, Hou H, Zhan C, Chen C, Lv W, Tao Q, Sun Z, Xia L. Correlation of Chest CT and RT-PCR Testing for Coronavirus Disease 2019 (COVID-19) in China: A Report of 1014 Cases. Ra- diology. 2020 Aug;296(2):E32-E40. doi: 10.1148/radiol.2020200642. Epub 2020 Feb 26. PMID: 32...

2019 doi
[16]

A systematic review of CT chest in COVID-19 diagnosis and its potential application in a surgical setting

Shao JM, Ayuso SA, Deerenberg EB, Elhage SA, Augenstein V A, Heniford BT. A systematic review of CT chest in COVID-19 diagnosis and its potential application in a surgical setting. Colorectal Dis. 2020 Sep;22(9):993-1001. doi: 10.1111/codi.15252. Epub 2020 Aug

2020 doi
[17]

PMID: 32644268; PMCID: PMC7361755
[18]

A Deep Learning Ap- proach for the Detection of COVID-19 from Chest X-Ray Images us- ing Convolutional Neural Networks

Saxena, Aditya, and Shamsheer Pal Singh. "A Deep Learning Ap- proach for the Detection of COVID-19 from Chest X-Ray Images us- ing Convolutional Neural Networks." arXiv preprint arXiv:2201.09952 (2022)

2022 arXiv
[19]

Comparative analysis of deep learning models for COVID- 19 detection

Kumari S, Ranjith E, Gujjar A, Narasimman S, Aadil Sha Zeelani HS. Comparative analysis of deep learning models for COVID- 19 detection. Global Transitions Proceedings. 2021;2(2):559-565. doi:10.1016/j.gltp.2021.08.030

2021 doi
[20]

Deep learning for COVID-19 detection based on CT images

Zhao, W., Jiang, W., Qiu, X. Deep learning for COVID-19 detection based on CT images. Sci Rep 11, 14353 (2021). https://doi.org/10.1038/s41598-021-93832-2

2021 doi
[21]

Deep learning for diagnosis of COVID- 19 using 3D CT scans

Serte S, Demirel H. Deep learning for diagnosis of COVID- 19 using 3D CT scans. Comput Biol Med. 2021;132:104306. doi:10.1016/j.compbiomed.2021.104306

2021
[22]

STAN-CT: Standardizing CT Image using Generative Adversarial Networks[J]

Selim M, Zhang J, Fei B, et al. STAN-CT: Standardizing CT Image using Generative Adversarial Networks[J]. AMIA Annu Symp Proc. 2020, 2020: 1100-1109

2020
[23]

COVID-19 Diagnosis from CT Images with Convolutional Neural Network Optimized by Marine Predator Optimization Algorithm

Huaping Jia, Junlong Zhao, Ali Arshaghi, "COVID-19 Diagnosis from CT Images with Convolutional Neural Network Optimized by Marine Predator Optimization Algorithm", BioMed Research International, vol. 2021, Article ID 5122962, 9 pages, 2021

2021
[24]

Learning imbalanced datasets with label-distribution-aware margin loss[J]

Cao K, Wei C, Gaidon A, et al. Learning imbalanced datasets with label-distribution-aware margin loss[J]. Advances in neural informa- tion processing systems, 2019, 32

2019
[25]

Calibrating deep neural net- works using focal loss[J]

Mukhoti J, Kulharia V , Sanyal A, et al. Calibrating deep neural net- works using focal loss[J]. Advances in Neural Information Processing Systems, 2020, 33: 15288-15299

2020
[26]

Class-balanced loss based on effective number of samples[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Cui Y , Jia M, Lin T Y , et al. Class-balanced loss based on effective number of samples[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2019: 9268-9277

2019

Pith tools

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

[1] [1]

Mosmeddata: Chest ct scans with covid-19 related findings dataset[J]

Morozov S P, Andreychenko A E, Pavlov N A, et al. Mosmeddata: Chest ct scans with covid-19 related findings dataset[J]. arXiv preprint arXiv:2005.06465, 2020

arXiv 2005

[2] [2]

Deep residual learning[J]

He K, Zhang X, Ren S, et al. Deep residual learning[J]. Image Recognition, 2015

work page 2015

[3] [3]

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

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

work page 2015

[4] [4]

Very deep convolutional networks for large-scale image recognition[J]

Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition[J]. arXiv preprint arXiv:1409.1556, 2014

arXiv 2014

[5] [8]

MisGAN: Learn- ing from Incomplete Data with Generative Adversarial Networks

Steven Cheng-Xian Li, Bo Jiang, Benjamin Marlin. "MisGAN: Learn- ing from Incomplete Data with Generative Adversarial Networks." International Conference on Learning Representations 2019

work page 2019

[6] [9]

Improved training of wasserstein gans

Gulrajani, Ishaan, et al. "Improved training of wasserstein gans." Advances in neural information processing systems 30 (2017)

2017

[7] [10]

Generative adversarial nets

Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y . "Generative adversarial nets." Advances in Neural Information Processing Systems. (2014)

work page 2014

[8] [11]

Deep generative image models using a laplacian pyramid of adversarial networks

Denton, E.L., Chintala, S., Fergus, R., et al. "Deep generative image models using a laplacian pyramid of adversarial networks." Advances in neural information processing systems. (2015)

work page 2015