Multi Source COVID-19 Detection via Kernel-Density-based Slice Sampling

Bo-Cheng Qiu; Chia-Ming Lee; Chih-Chung Hsu; Fang-Ying Lin; I-An Tsai; Jung-Tse Tsai; Ming-Han Sun; Ting-Yao Chen; Yu-Fan Lin

REVIEW 4 major objections 4 minor 20 references

Multi Source COVID-19 Detection via Kernel-Density-based Slice Sampling

T0 review · 4 major / 4 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 lightweight EfficientNet-B7 with kernel-density slice sampling reaches a 94.68 F1 score on multi-source COVID-19 CT classification, edging out a Swin Transformer while cutting slice redundancy by 70%.

desk verdict The headline F1 comparison doesn't hold up because the paper contradicts itself on Source 2's validation composition, making the pooled F1 ill-defined. read the letter →

arxiv 2507.01564 v2 pith:ALSQKIXW submitted 2025-07-02 eess.IV cs.CV

Chia-Ming Lee , Bo-Cheng Qiu , Ting-Yao Chen , Ming-Han Sun , Fang-Ying Lin , Jung-Tse Tsai , I-An Tsai , Yu-Fan Lin

show 1 more author

Chih-Chung Hsu

This is my paper · ORCID

classification eess.IVcs.CV

keywords multi-sourceCOVID-19detectionchestCTclassificationkernel-densityslicesamplingSpatial-SliceFeatureLearningEfficientNet-B7SwinTransformerdomainshiftdatasetbalance

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 is a challenge solution trying to show that a compact 2D CNN with careful slice selection can handle multi-source chest CT COVID-19 detection as well as or better than a heavier transformer. The authors build on the SSFL framework, add a kernel-density-based slice sampler (KDS) that picks eight representative slices per scan, and compare EfficientNet-B7 against Swin Transformer-Base under identical preprocessing. On the validation set, EfficientNet reaches an F1 of 94.68 versus 93.34 for Swin, with a 70% reduction in data redundancy. If correct, the result matters because it suggests domain variability across hospitals can be tamed by preprocessing and slice selection rather than by model scale alone.

What carries the argument

Kernel-Density-based Slice Sampling (KDS). It estimates the distribution of lung-area values across a CT volume with a Gaussian kernel density estimator using Scott's rule for bandwidth, builds the cumulative distribution function, partitions it into eight percentile intervals, and samples one slice per interval proportional to its contribution. This yields eight representative slices per scan. The machinery's job is to make slice selection adaptive to a scan's anatomy rather than fixed-position, preserving coverage while cutting redundancy by about 70%. The SSFL preprocessing stage, which combines lung region extraction, quality control, cropping, and resizing to 256 by 256 pixels, standardizes inputs across sites before KDS is applied.

What would settle it

Recompute the overall F1 on a class-balanced validation set, or inspect Source 2's confusion matrix: if Source 2 contains cases from only one class, its per-source F1 is vacuous, and dropping it or rebalancing it changes the EfficientNet-vs-Swin ranking; that would settle whether the 94.68 claim reflects true multi-source ability.

Watch

Extended reading notes

Core claim

The central claim is that the SSFL+KDS pipeline converts each full CT volume into eight informative slices, and that a lightweight 2D CNN trained on those slices generalizes across four medical centers. The paper reports EfficientNet-B7 beating Swin Transformer-Base on F1 (94.68 vs 93.34) and AUC-ROC (0.9813 vs 0.9797) on the challenge validation set, and attributes the difference to slice sampling that preserves anatomical coverage while removing redundant frames. It also argues that Source 2's degenerate validation subset, which contains only one class, distorts per-source metrics and shows why dataset balance is essential in multi-institutional evaluation.

Load-bearing premise

The load-bearing premise is that the validation F1 scores, computed on a validation set in which one site's subset is single-class, fairly measure cross-site generalization rather than class-imbalance artifacts.

Editorial extensions

If this is right

Deployments can use a lightweight EfficientNet-B7 rather than a large transformer for multi-source COVID-19 CT screening, with no loss and possibly a gain in F1.
KDS-style slice sampling can cut the number of slices fed to a model by about 70%, which reduces compute, memory, and reading time.
Because slice selection preserves anatomical coverage, the same preprocessing should transfer to other multi-center CT tasks and to scans with different lengths and spacing.
The paper's per-source results warn that an imbalanced validation subset can distort apparent model quality, so multi-institutional benchmarks should report class-balanced per-site metrics.

Reading between the lines

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

A natural extension is to apply KDS to other volumetric tasks, such as lung nodule detection or lesion load quantification, where slice redundancy is high and anatomy varies by scanner.
The 70% redundancy reduction implies substantial efficiency gains: training on eight slices instead of dozens could allow larger batch sizes, longer training, or lower-cost inference, but the paper does not measure those gains directly.
The headline comparison rests on validation-set labels; since the test set is unlabeled, true generalization beyond validation remains untested, and a balanced per-source re-evaluation would be the decisive check.
The paper's own emphasis on dataset balance suggests that a macro-F1 computed across balanced per-source subsets would be a more honest headline metric for cross-site robustness.

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

The headline finding—EfficientNet-B7 at 94.68 F1 vs Swin's 93.34—doesn't survive contact with the paper's own tables. Table 1 says Source 2's validation set has 0 COVID and 45 non-COVID scans; Section 3.3 and the Table 2 footnote say Source 2's validation set contains only COVID (positive) samples. Both cannot be true. With a single-class validation source, the pooled F1 is ill-defined: per-source F1 depends on whether the model always predicts the sole present class, not on diagnostic quality. The 1.34-point gap is therefore not a trustworthy cross-source comparison. The stress-test note is right on this.

Credit where it's due: the paper reports new empirical measurements on the PHAROS-AFE-AIMI validation set, the 70% data-reduction claim is practical, and comparing a lightweight CNN against a transformer under identical preprocessing is a legitimate exercise. The authors also correctly flag that class imbalance in Source 2's validation set matters. But the contradiction undercuts that same point.

The bigger methodological hole is the missing ablation. KDS is presented as the key contribution, but there's no comparison to random slice sampling or to a no-KDS baseline. The performance might come from the lung cropping, the quality control, or simply the choice of backbone. The paper also lacks error bars or fold-level results; training for three epochs with early stopping is thin, and the reported AUC values have no confidence intervals.

Novelty is modest: the SSFL and KDS components are the authors' own previously published work, so the new content is essentially a benchmark application. That's fine for a challenge report, but it shouldn't be oversold.

The paper is for readers working on COVID-19 CT detection and domain generalization. There's a useful data point buried here, but as submitted the central claim is not supported. I'd want the authors to fix the Source 2 contradiction, add an ablation, and report standard errors before I'd trust the 94.68 number.

Recommendation: if this comes to your desk, desk-reject with an invitation to resubmit after corrections; the current internal contradiction is too load-bearing to send to referees as is.

Referee Report

4 major / 4 minor

Summary. The paper describes a solution to the Multi-Source COVID-19 Detection Challenge, classifying chest CT scans from four medical centers. The proposed pipeline combines the authors' earlier Spatial-Slice Feature Learning (SSFL) framework with a Kernel-Density-based Slice Sampling (KDS) step that selects eight representative slices per scan after lung-region cropping. The authors compare EfficientNet-B7 and Swin Transformer on the challenge validation set, reporting an F1-score of 94.68% for EfficientNet versus 93.34% for Swin, and conclude that the KDS-based pipeline is effective for multi-source COVID-19 detection. The paper also reports per-source F1 and AUC values and discusses the impact of class imbalance in the validation set.

Significance. If the reported results were valid, the paper would provide a practically relevant demonstration that a lightweight 2D CNN with careful preprocessing can match or exceed a transformer on multi-source chest CT classification, and the idea of using kernel-density-based slice selection to handle variable scan lengths is interesting. The comparison on a public challenge benchmark is a useful reference point. However, the current evaluation is undermined by a direct contradiction about the composition of the Source 2 validation set, by the absence of any ablation of the KDS sampling component, and by the lack of uncertainty quantification. These issues must be resolved before the central claim of cross-source generalization can be accepted.

major comments (4)

[Section 3.3 / Table 1 / Table 2] Section 3.3 and the Table 2 footnote state that Source 2's validation set contains only positive (COVID) samples, but Table 1(a) lists 0 COVID and 45 non-COVID validation scans for Source 2. These statements are mutually exclusive. Because F1 is class-dependent, the pooled F1 scores in Table 3 are not well-defined as a multi-source measure if one source's validation fold is single-class; the per-source Source 2 values (100* for Swin and 49.21* for EfficientNet) illustrate the artifact. The headline comparison of 94.68 vs. 93.34 cannot serve as evidence of cross-source generalization until the composition is corrected and the metrics are recomputed on a balanced or per-class basis.
[§2.2 and Abstract] The KDS sampling is the paper's key methodological contribution and the Abstract attributes the results to 'our KDS-based pipeline,' yet no experiment replaces KDS with random slice selection, uniform sampling, or full-volume input. Without this ablation, the specific contribution of KDS to the reported F1 scores is unsupported, and the central claim that the KDS pipeline is effective is not established.
[§3.2 and Tables 2–3] Section 3.2 describes a stratified 5-fold cross-validation strategy, but Tables 2 and 3 report only point estimates with no fold-level results, standard deviations, or confidence intervals. With per-source validation sizes of 45–88 scans and a small overall margin (94.68 vs. 93.34), the reported differences may be within random variation; the superiority claim requires uncertainty quantification.
[Conclusion] The Conclusion states that the approach 'reduc[es] data redundancy by 70%,' but no calculation or table in the manuscript supports this figure. The claim should either be derived from the slice-count data in Table 1(b) or removed.

minor comments (4)

[Eq. (1)] Equation (1) defines a weighted average but does not specify the weights w(p,q); presumably they are uniform over the 3×3 window, but this should be stated.
[Eq. (3)] In Eq. (3), the symbol s is used for both the number of samples in the KDE sum and is described as 'smoothing factor,' while h is the bandwidth; the roles of s and h should be clarified.
[Tables 2 and 3] Tables 2 and 3 contain a typo: 'EfficentNet-B7' should be 'EfficientNet-B7'.
[Section 3.3] Section 3.3 states 'Both models show lower performance on Source 2,' but Table 2 reports a Swin Transformer F1 of 100* on Source 2; the text and table are inconsistent.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the reported F1 scores are empirical benchmark measurements, not derived outputs of the cited SSFL/KDS equations.

full rationale

The paper's central claims are measured F1/AUC values on the public COVID-19-CT-DB validation set, obtained by training EfficientNet-B7 and Swin Transformer on the challenge data. The preprocessing described in Equations (1)-(5) (minimum filtering, threshold-based lung cropping, and kernel-density slice sampling) is a standard, parameter-light procedure; Scott's rule sets the bandwidth, and the 8-slice sampling rule is an explicit method choice, not a fitted parameter. The paper does not define F1 in terms of the KDS equations, and it does not fit any parameter to the validation set's F1 and then call that a prediction. The extensive self-citations to the authors' SSFL/KDS prior work ([5]-[8]) are methodological provenance, not load-bearing evidence: no uniqueness theorem from those papers is invoked, and no ansatz is smuggled in as an external mathematical fact. The internal contradiction between Section 3.3 (Source 2 validation 'contains only positive samples') and Table 1 (Source 2 validation has 0 COVID and 45 non-COVID scans) is a real data-integrity and evaluation-validity concern, but it is not circularity; even if the validation fold is degenerate, the F1 numbers remain empirical outputs rather than outputs forced by the method's definitions. The absence of a random-slice ablation weakens the causal attribution to KDS, but a missing ablation is an experimental-design limitation, not a circular derivation. Accordingly, no circular step can be exhibited, and the honest finding is no significant circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The core pipeline uses standard statistical and deep-learning components. The main load-bearing assumptions are the hand-set segmentation threshold, the representativeness of KDE-based slice sampling, and the fairness of the validation set. No new physical or algorithmic entities are introduced.

free parameters (3)

Lung binarization threshold t
Empirically assigned in Section 2.1, Eq. (2), with no reported value. It affects segmentation of every scan and is not validated.
KDE smoothing factor s = 100
Section 2.2, Eq. (3) sets s=100 as a smoothing factor without justification; it appears in the normalization term of the KDE and is a hand-chosen constant.
Number of representative slices = 8
Section 2.2 and Section 3.2 fix eight slices per scan without a sensitivity analysis, so it is a design choice rather than a data-driven optimum.

assumptions (4)

standard math Kernel density estimation and the Gaussian kernel formula provide a valid density estimate of lung areas.
Section 2.2, Eqs. (3)-(4). These are standard statistical tools, but the normalization with s=100 is nonstandard and unexplained.
domain assumption Fixed-threshold binarization with empirically assigned threshold t correctly isolates lung regions across all scanners.
Section 2.1, Eq. (2). No threshold value is given and no segmentation quality check is reported, yet every downstream slice is cropped based on this mask.
domain assumption Slices sampled proportionally to the KDE of lung-area distribution preserve diagnostically relevant COVID-19 information.
Section 2.2. The paper asserts anatomical coverage, but offers no evidence that KDS-selected slices outperform random slices or retain lesions better than other sampling schemes.
domain assumption The validation set provides a fair, class-balanced evaluation across sources.
Section 3.1 Table 1 shows Source 2 validation has only one class, so this assumption is violated and undermines the headline F1.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Multi Source COVID-19 Detection via Kernel-Density-based Slice Sampling." pith.science (2026). https://pith.science/paper/ALSQKIXW

@misc{pith2026250701564,
  author       = {Pith},
  title        = {Pith review of: Multi Source COVID-19 Detection via Kernel-Density-based Slice Sampling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ALSQKIXW}},
  note         = {Machine review of arXiv:2507.01564}
}

read the original abstract

We present our solution for the Multi-Source COVID-19 Detection Challenge, which classifies chest CT scans from four distinct medical centers. To address multi-source variability, we employ the Spatial-Slice Feature Learning (SSFL) framework with Kernel-Density-based Slice Sampling (KDS). Our preprocessing pipeline combines lung region extraction, quality control, and adaptive slice sampling to select eight representative slices per scan. We compare EfficientNet and Swin Transformer architectures on the validation set. The EfficientNet model achieves an F1-score of 94.68%, compared to the Swin Transformer's 93.34%. The results demonstrate the effectiveness of our KDS-based pipeline on multi-source data and highlight the importance of dataset balance in multi-institutional medical imaging evaluation.

Figures

Figures reproduced from arXiv: 2507.01564 by the authors.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 12 canonical work pages

[1]

A large imaging database and novel deep neural ar- chitecture for covid-19 diagnosis

Anastasios Arsenos, Dimitrios Kollias, and Stefanos Kol- lias. A large imaging database and novel deep neural ar- chitecture for covid-19 diagnosis. In 2022 IEEE 14th Im- age, Video, and Multidimensional Signal Processing Work- shop (IVMSP), page 1–5. IEEE, 2022. 1

work page 2022
[2]

Data-driven covid-19 detection through medical imaging

Anastasios Arsenos, Andjoli Davidhi, Dimitrios Kollias, Panos Prassopoulos, and Stefanos Kollias. Data-driven covid-19 detection through medical imaging. In 2023 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW), page 1–5. IEEE, 2023. 1

work page 2023
[3]

Imagenet: A large-scale hierarchical image database

Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 3

2009
[4]

Covid- 19 computer-aided diagnosis through ai-assisted ct imaging analysis: Deploying a medical ai system

Demetris Gerogiannis, Anastasios Arsenos, Dimitrios Kol- lias, Dimitris Nikitopoulos, and Stefanos Kollias. Covid- 19 computer-aided diagnosis through ai-assisted ct imaging analysis: Deploying a medical ai system. In 2024 IEEE In- ternational Symposium on Biomedical Imaging (ISBI), pages 1–4. IEEE, 2024. 1

work page 2024
[5]

Strong baseline and bag of tricks for covid-19 detection of ct scans, 2023

Chih-Chung Hsu, Chih-Yu Jian, Chia-Ming Lee, Chi-Han Tsai, and Sheng-Chieh Dai. Strong baseline and bag of tricks for covid-19 detection of ct scans, 2023. 1

work page 2023
[6]

Bag of tricks of hybrid network for covid-19 detection of ct scans

Chih-Chung Hsu, Chih-Yu Jian, Chia-Ming Lee, Chi-Han Tsai, and Shen-Chieh Tai. Bag of tricks of hybrid network for covid-19 detection of ct scans. In 2023 IEEE Interna- tional Conference on Acoustics, Speech, and Signal Process- ing Workshops (ICASSPW), pages 1–4, 2023. 1

work page 2023
[7]

A closer look at spatial-slice features learning for covid-19 detection

Chih-Chung Hsu, Chia-Ming Lee, Yang Fan Chiang, Yi- Shiuan Chou, Chih-Yu Jiang, Shen-Chieh Tai, and Chi-Han Tsai. A closer look at spatial-slice features learning for covid-19 detection. In 2024 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition Workshops (CVPRW), pages 4924–4934, 2024. 1, 3

work page 2024
[8]

Simple 2d convolutional neural network-based ap- proach for covid-19 detection, 2024

Chih-Chung Hsu, Chia-Ming Lee, Yang Fan Chiang, Yi- Shiuan Chou, Chih-Yu Jiang, Shen-Chieh Tai, and Chi-Han Tsai. Simple 2d convolutional neural network-based ap- proach for covid-19 detection, 2024. 1

work page 2024

Show all 20 references

[9]

Deep neural archi- tectures for prediction in healthcare

Dimitrios Kollias, Athanasios Tagaris, Andreas Stafylopatis, Stefanos Kollias, and Georgios Tagaris. Deep neural archi- tectures for prediction in healthcare. Complex & Intelligent Systems, 4(2):119–131, 2018. 1

2018
[10]

Deep transparent prediction through latent repre- sentation analysis

Dimitrios Kollias, N Bouas, Y Vlaxos, V Brillakis, M Se- feris, Ilianna Kollia, Levon Sukissian, James Wingate, and S Kollias. Deep transparent prediction through latent repre- sentation analysis. arXiv preprint arXiv:2009.07044, 2020. 1

2009 arXiv
[11]

Transpar- ent adaptation in deep medical image diagnosis

Dimitris Kollias, Y Vlaxos, M Seferis, Ilianna Kollia, Levon Sukissian, James Wingate, and Stefanos D Kollias. Transpar- ent adaptation in deep medical image diagnosis. In TAILOR, page 251–267, 2020. 1

2020
[12]

Mia-cov19d: Covid-19 detection through 3-d chest ct image analysis

Dimitrios Kollias, Anastasios Arsenos, Levon Soukissian, and Stefanos Kollias. Mia-cov19d: Covid-19 detection through 3-d chest ct image analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision , page 537–544, 2021. 1

2021
[13]

Ai-mia: Covid-19 detection and severity analysis through medical imaging

Dimitrios Kollias, Anastasios Arsenos, and Stefanos Kollias. Ai-mia: Covid-19 detection and severity analysis through medical imaging. In European Conference on Computer Vi- sion, page 677–690. Springer, 2022. 1

2022
[14]

Ai-enabled analysis of 3-d ct scans for diagnosis of covid-19 & its severity

Dimitrios Kollias, Anastasios Arsenos, and Stefanos Kollias. Ai-enabled analysis of 3-d ct scans for diagnosis of covid-19 & its severity. In 2023 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICAS- SPW), page 1–5. IEEE, 2023. 1

2023
[15]

A deep neural architecture for harmonizing 3-d input data analysis and decision making in medical imaging

Dimitrios Kollias, Anastasios Arsenos, and Stefanos Kollias. A deep neural architecture for harmonizing 3-d input data analysis and decision making in medical imaging. Neuro- computing, 542:126244, 2023. 1

2023
[16]

Domain adaptation explainability & fairness in ai for medical image analysis: Diagnosis of covid-19 based on 3-d chest ct-scans

Dimitrios Kollias, Anastasios Arsenos, and Stefanos Kol- lias. Domain adaptation explainability & fairness in ai for medical image analysis: Diagnosis of covid-19 based on 3-d chest ct-scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition ...

2024
[17]

Sam2clip2sam: Vision language model for segmentation of 3d ct scans for covid-19 detection

Dimitrios Kollias, Anastasios Arsenos, James Wingate, and Stefanos Kollias. Sam2clip2sam: Vision language model for segmentation of 3d ct scans for covid-19 detection. arXiv preprint arXiv:2407.15728, 2024. 1

2024 arXiv
[18]

Swin Transformer: Hierarchical Vision Transformer using Shifted Windows

Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical Vision Transformer using Shifted Windows . In 2021 IEEE/CVF International Conference on Computer Vi- sion (ICCV), pages 9992–10002, Los Alamitos, CA, USA,

2021
[19]

Efficientnet: Rethinking model scaling for convolutional neural networks

Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proceedings of the International conference on machine learning (ICML) , pages 6105–6114, 2019. 3

2019
[2021]

IEEE Computer Society. 3

Pith tools

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

[1] [1]

A large imaging database and novel deep neural ar- chitecture for covid-19 diagnosis

Anastasios Arsenos, Dimitrios Kollias, and Stefanos Kol- lias. A large imaging database and novel deep neural ar- chitecture for covid-19 diagnosis. In 2022 IEEE 14th Im- age, Video, and Multidimensional Signal Processing Work- shop (IVMSP), page 1–5. IEEE, 2022. 1

work page 2022

[2] [2]

Data-driven covid-19 detection through medical imaging

Anastasios Arsenos, Andjoli Davidhi, Dimitrios Kollias, Panos Prassopoulos, and Stefanos Kollias. Data-driven covid-19 detection through medical imaging. In 2023 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW), page 1–5. IEEE, 2023. 1

work page 2023

[3] [3]

Imagenet: A large-scale hierarchical image database

Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 3

2009

[4] [4]

Covid- 19 computer-aided diagnosis through ai-assisted ct imaging analysis: Deploying a medical ai system

Demetris Gerogiannis, Anastasios Arsenos, Dimitrios Kol- lias, Dimitris Nikitopoulos, and Stefanos Kollias. Covid- 19 computer-aided diagnosis through ai-assisted ct imaging analysis: Deploying a medical ai system. In 2024 IEEE In- ternational Symposium on Biomedical Imaging (ISBI), pages 1–4. IEEE, 2024. 1

work page 2024

[5] [5]

Strong baseline and bag of tricks for covid-19 detection of ct scans, 2023

Chih-Chung Hsu, Chih-Yu Jian, Chia-Ming Lee, Chi-Han Tsai, and Sheng-Chieh Dai. Strong baseline and bag of tricks for covid-19 detection of ct scans, 2023. 1

work page 2023

[6] [6]

Bag of tricks of hybrid network for covid-19 detection of ct scans

Chih-Chung Hsu, Chih-Yu Jian, Chia-Ming Lee, Chi-Han Tsai, and Shen-Chieh Tai. Bag of tricks of hybrid network for covid-19 detection of ct scans. In 2023 IEEE Interna- tional Conference on Acoustics, Speech, and Signal Process- ing Workshops (ICASSPW), pages 1–4, 2023. 1

work page 2023

[7] [7]

A closer look at spatial-slice features learning for covid-19 detection

Chih-Chung Hsu, Chia-Ming Lee, Yang Fan Chiang, Yi- Shiuan Chou, Chih-Yu Jiang, Shen-Chieh Tai, and Chi-Han Tsai. A closer look at spatial-slice features learning for covid-19 detection. In 2024 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition Workshops (CVPRW), pages 4924–4934, 2024. 1, 3

work page 2024

[8] [8]

Simple 2d convolutional neural network-based ap- proach for covid-19 detection, 2024

Chih-Chung Hsu, Chia-Ming Lee, Yang Fan Chiang, Yi- Shiuan Chou, Chih-Yu Jiang, Shen-Chieh Tai, and Chi-Han Tsai. Simple 2d convolutional neural network-based ap- proach for covid-19 detection, 2024. 1

work page 2024