Pith. sign in

REVIEW 6 major objections 5 minor 43 references

Taming Domain Shift in Multi-source CT-Scan Classification via Input-Space Standardization

T0 review · 6 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 This paper claims that standardizing CT inputs before training—cropping to lung anatomy and sampling slices by density—cuts inter-source feature variance by 75% and lifts COVID-19 classification F1 from 70.7% to 94.7%.

desk verdict Useful post-hoc analysis of the authors' own CT preprocessing pipeline; the empirical gains look real, but the anatomical-alignment mechanism is oversold. read the letter →

arxiv 2507.19858 v1 pith:WD45EXJ2 submitted 2025-07-26 eess.IV cs.CEcs.CVcs.LG

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.CEcs.CVcs.LG
keywords domainshiftinput-spacestandardizationCOVID-19CTclassificationSSFL++KDSslicesamplinginter-sourcevariancemulti-sourcegeneralizationpreprocessingpipeline
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 argues that domain shift between CT scanners can be reduced at the input level, before any model trains, rather than fixed afterward. It uses a two-step preprocessing pipeline: SSFL++ crops each scan to a consistent lung-centric bounding box, and KDS picks eight slices at fixed percentiles of the lung-density distribution so that every scan covers the same anatomical range. The paper reports that this standardization lowers inter-source feature variance from 3.49 to 0.87, a 75% reduction, and raises COVID-19 classification F1 from 70.73% to 94.68% on EfficientNet-B3 and from 68.53% to 93.34% on Swin Transformer. Because the gains appear across CNN and Transformer architectures, the paper concludes the improvement comes from the preprocessing itself, not from a particular model. A reader should care because input-space standardization needs no domain labels, is computationally cheap, and could transfer to other multi-institutional imaging tasks.

What carries the argument

The load-bearing machinery is the two-stage preprocessing pipeline. SSFL++ (Spatial-Slice Feature Learning) performs spatial standardization: it applies morphological filtering and adaptive binarization to isolate lung regions, then crops every scan to the minimal bounding box covering the union of lung masks, removing source-specific field-of-view and background variation. KDS (Kernel-Density-based Slice Sampling) performs temporal standardization: it estimates the lung-tissue area distribution across slices via kernel density estimation and samples eight slices at fixed percentiles (5%, 15%, ..., 95%) of that distribution, so each volume yields the same anatomical coverage from apex to base. The analysis treats these two modules as complementary: SSFL++ establishes a consistent anatomical frame, and KDS exploits that frame to align temporal sampling; the paper's quantitative evidence for the mechanism is the drop in inter-source variance and the ablation pattern across architectures.

What would settle it

Compute the true anatomical level (for example, distance from the lung apex or the carina) of the eight KDS-selected slices in each source; if the percentile-selected slices vary in anatomical position across sources more than uniformly sampled slices do, the claimed temporal consistency is falsified. A second check: apply the pipeline to CTs with severe fibrosis or pleural effusion and measure whether inter-source variance still drops by roughly 75%.

Watch

Extended reading notes

Core claim

The paper's central claim is that a preventive input-space standardization framework—spatial alignment through SSFL++ combined with temporal alignment through KDS—maps disparate multi-source CT scans into a consistent target space, and that this mapping is the cause of the large cross-source accuracy gains. Quantitatively, the complete pipeline reduces inter-source variance from 3.49 to 0.87 (a 75% drop) while improving macro F1 for COVID-19 classification from 70.73% to 94.68% with EfficientNet-B3 and from 68.53% to 93.34% with Swin Transformer. The feature-space analysis shows a trade-off: spatial-only standardization increases local class separability but leaves source-specific clustering intact, and only the addition of KDS substantially improves cross-source consistency. The authors interpret this as evidence that temporal heterogeneity—inconsistent slice selection—is the dominant component of domain shift in CT classification, and that standardizing it at the input is an effective inductive bias.

Load-bearing premise

The whole argument rests on KDS's assumption that fixed percentiles of the lung-tissue-area density curve mark the same anatomical positions from scan to scan; if lung deformation, field-of-view, or reconstruction changes the density profile, the selected slices will not be anatomically aligned, and the temporal standardization that drives the variance reduction would fail.

Editorial extensions

If this is right

  • Because the gain is attributed to preprocessing rather than architecture, the SSFL++ and KDS pipeline can be dropped into any CT classifier, including future backbones, without retraining domain-adversarial components.
  • The reported 75% inter-source variance reduction gives a concrete, measurable target: other domain-robustness methods can be compared by how much they reduce this variance, not only by end accuracy.
  • Temporal standardization being the dominant factor suggests that scan-length and slice-selection protocols deserve as much attention as scanner hardware when building multi-institutional datasets.
  • If the mechanism holds, the same idea should improve other volumetric medical imaging tasks where anatomical density distributions are informative, such as MRI or follow-up CT comparison.
  • The competition first place (77.6% F1 on a hidden test) indicates the pipeline transfers to an unseen evaluation set, though with a gap from internal validation.

Reading between the lines

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

  • A direct consequence the paper leaves implicit: KDS's fixed-percentile sampling may misalign slices when pathology distorts the lung-density distribution (e.g., fibrosis, collapse, large effusions), so the pipeline's robustness is likely bounded by how anatomical the density profile remains; this is testable on pathological cohorts.
  • The separability-versus-consistency trade-off suggests a tuning knob the paper does not explore: changing the percentile grid (e.g., skipping the most diseased apex/base slices) could rebalance local accuracy and cross-source generalization for specific deployment sites.
  • The architecture-independence claim, if general, implies that input-space standardization could be combined with post-hoc methods like VREx; the paper does not test whether their benefits are additive, but the different error sources make such a combination plausible.
  • The hidden-test gap (77.6% vs. 94.68% internal F1) hints at unmodeled domain factors beyond spatial framing and slice selection, such as reconstruction kernels or contrast phase, which the current preprocessing does not standardize.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

6 major / 5 minor

Summary. This paper analyzes the SSFL++/KDS input-space preprocessing pipeline for multi-source COVID-19 CT classification. The authors argue that spatial standardization (lung-centric cropping) and temporal standardization (Kernel-Density-based slice sampling at fixed percentiles of the lung-area distribution) reduce inter-source feature variance by 75% and improve macro F1 from 70.73% to 94.68% (EfficientNet-B3) and from 68.53% to 93.34% (Swin Transformer), with first place in the PHAROS-AFE-AIMI competition. Feature-space metrics (Fisher score, separability, inter-source variance) are used to explain the trade-off between local discriminability and cross-source consistency.

Significance. The work is potentially valuable because it offers a mechanistic, feature-space account of a preprocessing-only solution to domain shift, with large and consistent gains across two architectures and an external competition result. The authors are explicit that the pipeline is from their prior work [16,17], and the central empirical claim is supported by internal ablations rather than by a derivation assuming the conclusion. However, the manuscript lacks error bars, defines one reported metric incompletely, and does not compare against any post-hoc domain adaptation baseline, so the significance of the mechanistic explanation is currently limited.

major comments (6)
  1. [Table 1 and Section 4.2] The metric 'Domain Consistency' is reported in Table 1 but never defined in Section 4.2; without its formula the table's cross-source consistency claim cannot be checked. The table also appears internally inconsistent with the text: if 'w/o KDS' denotes SSFL++ only (as the surrounding text suggests), its inter-source variance (3.49) and domain consistency (1.91) are far worse than the baseline (1.06 and 0.22), yet Section 4.3 states that SSFL++ alone provides a modest improvement and is a necessary foundation. Clarify the method labels and reconcile the numbers.
  2. [Section 3.3, Eq. (2)] The mechanistic claim that KDS provides 'consistent anatomical coverage of lung anatomy from apex to base' requires that fixed percentiles of the lung-tissue-area density distribution correspond to the same anatomical z-positions across scans. This premise is not established: scans with different z-coverage, pathology-induced density shifts, or reconstruction differences will map the same percentile to different anatomical levels. Please provide per-source evidence for the anatomical alignment of the selected slices (e.g., histograms of selected slice positions or comparisons against anatomical landmarks), or revise the stated mechanism.
  3. [Section 3.3, Eq. (2)] The percentile set p ∈ {0.05, 0.15, ..., 0.85, 0.95} contains 10 values, not 8 as stated in the text and in Eq. (3)'s aggregation over 8 slices. This is a concrete mismatch in the definition of the core preprocessing method; please correct the count or the percentile set.
  4. [Section 4, Figure 4, and Table 3] The paper concludes that temporal heterogeneity (KDS) is the more critical component and that 'KDS alone has minimal impact,' but Table 3 contains no KDS-only condition: the rows are baseline, SSFL++ only, and SSFL++ plus KDS. Without a KDS-only F1/AUC measurement, the relative contribution of KDS and the claimed synergy cannot be assessed from the reported experiments.
  5. [Section 5.3, Table 3] All results are said to come from a 5-fold cross-validation, yet no standard deviations, confidence intervals, or per-fold values are reported. Given that the central quantitative claims are large F1 gaps and a 75% variance reduction (3.49 to 0.87), the stability of these numbers across folds is essential; please report mean ± std (or per-fold results) for at least the main ablations and Table 1.
  6. [Section 4.1 and Introduction] The paper frames the contribution as showing that input-space standardization is 'more effective than correcting for domain shift post-hoc,' but no experiment compares against any post-hoc domain adaptation method (e.g., domain-adversarial training, IRM, or VREx). The broad comparative claim is therefore unsupported; either include such baselines on the same data or restrict the claim to a demonstration that the pipeline is effective.
minor comments (5)
  1. [Figure 5] The bottom label 'Baseline (no crop)' is inconsistent with the 'w/o KDE' notation used elsewhere in the same figure; unify the method names for clarity.
  2. [Table 5] The AUC for Source 2 is marked with an asterisk but no value is given; either report the value or explain explicitly why it is omitted.
  3. [Section 5.6] The sentence 'The outcomes are consistent with 7' should read 'with Figure 7' and the figure should be referenced explicitly in the text.
  4. [References] Reference [8] is formatted as 'Hochreiter et al.' rather than as a proper citation with the full author list; please correct it.
  5. [Abstract and Section 4.3] The abstract's phrase 'proving the benefits stem from the preprocessing itself' is too strong given the absence of statistical tests and post-hoc baselines; consider 'consistent with' or 'support' instead.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: measured ablations and external competition support the central claims; KDS anatomical premise is an assumption, not a derivation.

full rationale

The paper's central claims are empirical measurements, not derivations from the cited pipeline. The 75% inter-source variance reduction (Table 1: 3.49 vs 0.87) and F1 gains (Table 3: 70.73 to 94.68 for EfficientNet-B3; 68.53 to 93.34 for Swin) come from controlled ablations (baseline, SSFL++ only, SSFL+++KDS) on the COVID-19-CT-DB dataset, and the pipeline is independently validated by a first-place PHAROS-AFE-AIMI competition result on a hidden test set. The paper adopts SSFL++/KDS from the authors' prior work [16,17], but it does not ask the cited papers to certify its quantitative findings; the variance and performance numbers are computed in this study. The only load-bearing premise imported from prior work is the assertion that KDS's fixed-percentile slice selection corresponds to consistent anatomical coverage (Section 3.3). That premise is an unsupported assumption and a genuine correctness risk, since variable scan coverage or pathology can break the percentile-to-anatomy mapping, but it is not a circularity: the paper does not define 'anatomical coverage' as the output of KDS, nor does it fit any parameter so that the conclusion follows by construction. No equation in the paper equates the claimed result to a fitted input. Self-citation is present but not load-bearing for the quantitative claims, so the circularity score is low.

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

The central claim depends on two hand-chosen preprocessing hyperparameters (KDS percentile set and slice count) and on three domain assumptions: anatomical alignment via lung-area percentiles, validity of the feature-space metrics as generalization proxies, and stability of single-run metric comparisons. No new entities are postulated.

free parameters (2)
  • KDS percentile set = {0.05, 0.15, ..., 0.85, 0.95} (ten percentiles in equation, eight slices in text)
    Hand-chosen set of slice-selection percentiles with no sensitivity analysis; sampling strategy directly determines the input representation and performance.
  • Number of sampled slices = 8
    The number of slices per scan is fixed to 8 by the percentile selection; no ablation on slice count is reported, yet it is a core design choice.
assumptions (3)
  • domain assumption Fixed percentiles of lung-tissue-area density correspond to the same anatomical positions across scans and sources.
    Section 3.3 assumes KDS percentile sampling ensures representative anatomical coverage; if lung area is distorted by pathology or acquisition, selected slices may not align anatomically.
  • domain assumption Feature-space metrics (Fisher score, separability, inter-source variance) computed from a trained EfficientNet-B3 are valid and stable proxies for generalization.
    Section 4.2 defines these metrics as measuring discriminability and cross-source consistency, but the paper does not validate their correlation with held-out performance, and no error bars are given.
  • domain assumption Differences in Table 1 metric values across configurations reflect preprocessing effects rather than random training variability.
    Single runs are reported without variance, so the 75% variance reduction and other metric changes are assumed to be reproducible across training seeds.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Taming Domain Shift in Multi-source CT-Scan Classification via Input-Space Standardization." pith.science (2026). https://pith.science/paper/WD45EXJ2

@misc{pith2026250719858,
  author       = {Pith},
  title        = {Pith review of: Taming Domain Shift in Multi-source CT-Scan Classification via Input-Space Standardization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WD45EXJ2}},
  note         = {Machine review of arXiv:2507.19858}
}
read the original abstract

Multi-source CT-scan classification suffers from domain shifts that impair cross-source generalization. While preprocessing pipelines combining Spatial-Slice Feature Learning (SSFL++) and Kernel-Density-based Slice Sampling (KDS) have shown empirical success, the mechanisms underlying their domain robustness remain underexplored. This study analyzes how this input-space standardization manages the trade-off between local discriminability and cross-source generalization. The SSFL++ and KDS pipeline performs spatial and temporal standardization to reduce inter-source variance, effectively mapping disparate inputs into a consistent target space. This preemptive alignment mitigates domain shift and simplifies the learning task for network optimization. Experimental validation demonstrates consistent improvements across architectures, proving the benefits stem from the preprocessing itself. The approach's effectiveness was validated by securing first place in a competitive challenge, supporting input-space standardization as a robust and practical solution for multi-institutional medical imaging.

Figures

Figures reproduced from arXiv: 2507.19858 by the authors.

Figure 1
Figure 1. Overview of the input-space standardization pipeline [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 3
Figure 3. Synergistic SSFL++ and KDS preprocessing visualiza [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Pipeline component analysis and cumulative effects. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figures from the paper (2 more)
Figure 6
Figure 6. Figure 6: Source clustering analysis through PCA visualization of learned feature representations across four sources under different [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Inter-source feature distance matrix. to class imbalance, though performance remains stable be￾tween validation and test sets. The outcomes are consistent with 7; predicting becomes increasingly challenging when the feature distance from other sources is greater. 5.7. …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 38 canonical work pages

  1. [1]

    Deep learning for the harmonization of structural mri scans: a survey

    Soolmaz Abbasi, Haoyu Lan, Jeiran Choupan, and Nasim Bahaei. Deep learning for the harmonization of structural mri scans: a survey. Biomed Eng Online, 23(1):90, 2024. 2

    work page 2024

  2. [2]

    Multi-modal medical image classi- fication using deep residual network and genetic algorithm

    Muhammad Haris Abid, Rehan Ashraf, Toqeer Mahmood, and CM Nadeem Faisal. Multi-modal medical image classi- fication using deep residual network and genetic algorithm. Plos one, 18(6):e0287786, 2023. 2

    work page 2023

  3. [3]

    Invariant risk minimization, 2020

    Martin Arjovsky, L ´eon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization, 2020. 1, 2

    work page 2020

  4. [4]

    Multistage transfer learning for medical images

    Gelan Ayana, Kokeb Dese, Ahmed Mohammed Abagaro, Kwangcheol Casey Jeong, Soon-Do Yoon, and Se-woon Choe. Multistage transfer learning for medical images. Arti- ficial Intelligence Review, 57(9):232, 2024. 2

    work page 2024

  5. [5]

    Adaptive distribution learning with statistical hypothesis testing for covid-19 ct scan classification

    Guan-Lin Chen, Chih-Chung Hsu, and Mei-Hsuan Wu. Adaptive distribution learning with statistical hypothesis testing for covid-19 ct scan classification. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 471–479, 2021. 2

    work page 2021

  6. [6]

    Advancing medical imaging informatics by deep learning-based domain adaptation

    Anirudh Choudhary, Li Tong, Yuanda Zhu, and May D Wang. Advancing medical imaging informatics by deep learning-based domain adaptation. Yearbook of medical in- formatics, 29(01):129–138, 2020. 1, 2

    work page 2020

  7. [7]

    An image is worth 16x16 words: Trans- formers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2

    arXiv 2010

  8. [8]

    Long short-term memory

    Hochreiter et al. Long short-term memory. Neural computa- tion, 9(8):1735–1780, 1997. 2

    work page 1997

Show all 43 references
  1. [9]

    Domain-adversarial train- ing of neural networks, 2016

    Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pas- cal Germain, Hugo Larochelle, Franc ¸ois Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial train- ing of neural networks, 2016. 1, 2

    2016

  2. [10]

    Domain adaptive video segmentation via temporal consis- tency regularization

    Dayan Guan, Jiaxing Huang, Aoran Xiao, and Shijian Lu. Domain adaptive video segmentation via temporal consis- tency regularization. In Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 8053–8064,

  3. [11]

    Domainatm: Domain adapta- tion toolbox for medical data analysis, 2022

    Hao Guan and Mingxia Liu. Domainatm: Domain adapta- tion toolbox for medical data analysis, 2022. 1, 2

    2022

  4. [12]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 7

    2016

  5. [13]

    Cmc-cov19d: Contrastive mixup classification for covid-19 diagnosis

    Junlin Hou, Jilan Xu, Rui Feng, Yuejie Zhang, Fei Shan, and Weiya Shi. Cmc-cov19d: Contrastive mixup classification for covid-19 diagnosis. In 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 454–461, 2021. 2

    2021

  6. [14]

    Cmc v2: Towards more ac- curate covid-19 detection with discriminative video priors,

    Junlin Hou, Jilan Xu, Nan Zhang, Yi Wang, Yuejie Zhang, Xiaobo Zhang, and Rui Feng. Cmc v2: Towards more ac- curate covid-19 detection with discriminative video priors,

  7. [15]

    Spatiotemporal feature learning based on two-step lstm and transformer for ct scans.arXiv preprint arXiv:2207.01579, 2022

    Chih-Chung Hsu, Chi-Han Tsai, Guan-Lin Chen, Sin-Di Ma, and Shen-Chieh Tai. Spatiotemporal feature learning based on two-step lstm and transformer for ct scans.arXiv preprint arXiv:2207.01579, 2022. 2

    2022 arXiv

  8. [16]

    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. IEEE, 2023. 1, 2

    2023

  9. [17]

    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 Confer- ence on Computer Vision and Pattern Recognition Work- shops (CVP...

    2024

  10. [18]

    Drct: Saving image super-resolution away from information bottleneck

    Chih-Chung Hsu, Chia-Ming Lee, and Yi-Shiuan Chou. Drct: Saving image super-resolution away from information bottleneck. In 2024 IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition Workshops (CVPRW) , pages 6133–6142, 2024. 2

    2024

  11. [19]

    Revisiting vision- language features adaptation and inconsistency for so- cial media popularity prediction

    Chih-Chung Hsu, Chia-Ming Lee, Yu-Fan Lin, Yi-Shiuan Chou, Chih-Yu Jian, and Chi-Han Tsai. Revisiting vision- language features adaptation and inconsistency for so- cial media popularity prediction. In Proceedings of the 32nd ACM International Conference on Multimedia , page 1...

    2024

  12. [20]

    A lifelong learning approach to brain mr segmentation across scanners and protocols, 2018

    Neerav Karani, Krishna Chaitanya, Christian Baumgartner, and Ender Konukoglu. A lifelong learning approach to brain mr segmentation across scanners and protocols, 2018. 2

    2018

  13. [21]

    Deep transparent prediction through latent represen- tation 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 represen- tation analysis. arXiv preprint arXiv:2009.07044, 2020. 1, 2

    2009 arXiv

  14. [22]

    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

  15. [23]

    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, 2

    2021

  16. [24]

    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, 2

    2022

  17. [25]

    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

  18. [26]

    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

  19. [27]

    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

  20. [28]

    Out-of-distribution generaliza- tion via risk extrapolation (rex), 2021

    David Krueger, Ethan Caballero, Joern-Henrik Jacobsen, Amy Zhang, Jonathan Binas, Dinghuai Zhang, Remi Le Priol, and Aaron Courville. Out-of-distribution generaliza- tion via risk extrapolation (rex), 2021. 1, 2

    2021

  21. [29]

    Re- thinking domain adaptation for machine learning over clini- cal language

    Egoitz Laparra, Steven Bethard, and Timothy A Miller. Re- thinking domain adaptation for machine learning over clini- cal language. JAMIA open, 3(2):146–150, 2020. 1, 2

    2020

  22. [30]

    Prompthsi: Universal hyperspectral image restoration with vision-language modulated frequency adaptation, 2025

    Chia-Ming Lee, Ching-Heng Cheng, Yu-Fan Lin, Yi-Ching Cheng, Wo-Ting Liao, Fu-En Yang, Yu-Chiang Frank Wang, and Chih-Chung Hsu. Prompthsi: Universal hyperspectral image restoration with vision-language modulated frequency adaptation, 2025. 2

    2025

  23. [31]

    Enhancing and adapt- ing in the clinic: Source-free unsupervised domain adapta- tion for medical image enhancement

    Heng Li, Ziqin Lin, Zhongxi Qiu, Zinan Li, Ke Niu, Na Guo, Huazhu Fu, Yan Hu, and Jiang Liu. Enhancing and adapt- ing in the clinic: Source-free unsupervised domain adapta- tion for medical image enhancement. IEEE Transactions on Medical Imaging, 43(4):1323–1336, 2024. 1, 2

    2024

  24. [32]

    Brooks, Nathan Jacobs, and Jin Chen

    Gongbo Liang, Sajjad Fouladvand, Jie Zhang, Michael A. Brooks, Nathan Jacobs, and Jin Chen. Ganai: Standardizing ct images using generative adversarial network with alterna- tive improvement. In 2019 IEEE International Conference on Healthcare Informatics (ICHI), pages 1–11, 2019. 2

    2019

  25. [33]

    Divide and conquer: Grounding a bleeding areas in gastrointestinal image with two-stage model, 2024

    Yu-Fan Lin, Bo-Cheng Qiu, Chia-Ming Lee, and Chih- Chung Hsu. Divide and conquer: Grounding a bleeding areas in gastrointestinal image with two-stage model, 2024. 2

    2024

  26. [34]

    Densesr: Image shadow removal as dense prediction, 2025

    Yu-Fan Lin, Chia-Ming Lee, and Chih-Chung Hsu. Densesr: Image shadow removal as dense prediction, 2025. 2

    2025

  27. [35]

    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

  28. [36]

    Stan-ct: Standardizing ct image using generative adversarial network, 2020

    Md Selim, Jie Zhang, Baowei Fei, Guo-Qiang Zhang, and Jin Chen. Stan-ct: Standardizing ct image using generative adversarial network, 2020. 2

    2020

  29. [37]

    Ct image harmonization for enhancing radiomics studies

    Md Selim, Jie Zhang, Baowei Fei, Guo-Qiang Zhang, and Jin Chen. Ct image harmonization for enhancing radiomics studies. In 2021 IEEE International Conference on Bioinfor- matics and Biomedicine (BIBM), pages 1057–1062, 2021. 2

    2021

  30. [38]

    Cross-vendor ct image data harmonization using cvh-ct, 2021

    Md Selim, Jie Zhang, Baowei Fei, Guo-Qiang Zhang, Gary Yeeming Ge, and Jin Chen. Cross-vendor ct image data harmonization using cvh-ct, 2021. 2

    2021

  31. [39]

    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, 4, 6, 7

    2019

  32. [40]

    Tent: Fully test-time adaptation by entropy minimization

    Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Ol- shausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726,

    2006 arXiv

  33. [41]

    Tent: Fully test-time adaptation by entropy minimization

    Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Ol- shausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021. 1

    2021

  34. [42]

    Multi-source covid-19 de- tection via variance risk extrapolation, 2025

    Runtian Yuan, Qingqiu Li, Junlin Hou, Jilan Xu, Yuejie Zhang, Rui Feng, and Hao Chen. Multi-source covid-19 de- tection via variance risk extrapolation, 2025. 1, 2

    2025

  35. [2021]

    IEEE Computer Society. 6, 7

Pith tools

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