Pith. sign in

REVIEW 5 major objections 8 minor 1 cited by

CycleGuardian: A Framework for Automatic RespiratorySound classification Based on Improved Deep clustering and Contrastive Learning

T0 review · 5 major / 8 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read CycleGuardian claims the best no-pretrained four-class result on the ICBHI2017 official split, with a 38M-parameter model that also deploys on Android.

desk verdict A genuine method contribution undermined by a headline score chosen on the evaluation split, no error bars, and mismatched per-class results; worth reviewing but needs major revision. read the letter →

arxiv 2502.00734 v1 pith:65VHFWVL submitted 2025-02-02 cs.SD cs.AIeess.AS

Yun Chu , Qiuhao Wang , Enze Zhou , Ling Fu , Qian Liu , Gang Zheng This is my paper · ORCID
classification cs.SDcs.AIeess.AS
keywords respiratorysoundclassificationICBHI2017deepclusteringcontrastivelearninggroup-mixspectrogramgroupinglightweightnetworkauscultation
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

CycleGuardian is a lightweight neural network that sorts respiratory sounds into four classes — normal, crackle, wheeze, and both — and the paper claims it achieves the best current four-class result on the official ICBHI2017 split among methods trained without pretrained weights: 82.06% specificity, 44.47% sensitivity, a score of 63.26%, and a 38M-parameter model. The result matters because the strongest existing systems use roughly 380MB transformer models pretrained on large image and audio datasets, which are impractical on phones and embedded stethoscopes. The paper's proposal is that grouping the spectrogram into short frame-level chunks, clustering those chunks, and contrastively separating samples can substitute for model scale and pretraining. A reader should care because it suggests accurate automated auscultation may be achievable in a deployable, mobile form.

What carries the argument

The load-bearing object is the IDEC-CPF pairing inside CycleGuardian. IDEC is an Improved Deep Embedding Clustering module that soft-assigns each group embedding to one of five clusters (intended as normal, noise, crackle, wheeze, and both) using a Student-t distribution, and CPF is a Cluster Projection Fusion module that maps the group features of each cluster into a single cluster feature and fuses the five cluster features into a global vector. A soft-cosine-similarity loss pushes the five cluster features apart, which the paper argues separates abnormal from normal and different abnormal types from each other. The group-mix contrastive branch supplies the complementary pressure: it replaces a fraction $1-\lambda$ of a sample's groups with groups from another sample and trains the global feature of the mixed sample to keep similarity $\lambda$ with the original and $1-\lambda$ with the donor.

What would settle it

Cross-tabulate the hard cluster assignment of every group against the known class of its breath cycle on the official validation split; if the five clusters do not align with the intended semantic categories (for instance, an adjusted Rand index near zero), then the clustering mechanism is not carrying the reported performance gain.

Watch

Extended reading notes

Core claim

The paper's central claim is that a small convolutional network, trained from scratch, can outperform much larger and pretrained systems on four-class respiratory sound classification when its features are organized by two complementary pressures: deep clustering with a soft-cosine constraint and group-mix contrastive learning. On the official ICBHI2017 split, CycleGuardian reports 82.06% specificity and 44.47% sensitivity (score 63.26%), which is roughly seven points above the best no-pretrained baseline and about one point above the best pretrained transformer baseline, at one tenth of that model's size. The method cuts each spectrogram into overlapping groups of 20 frames, embeds each group, and routes the embeddings through two branches: an improved deep embedding clustering branch that soft-assigns groups to five clusters and fuses them, and a contrastive branch that mixes groups from different samples and trains the mixed global feature to sit proportionally close to both sources.

Load-bearing premise

The five clusters learned by the deep clustering module are assumed to correspond to the semantic categories normal, noise, crackle, wheeze, and both, and the paper never verifies that the clusters actually align with those categories.

Editorial extensions

If this is right

  • A 38M-parameter respiratory sound classifier can outperform 380MB transformer models on the standard four-class benchmark without any pretrained weights.
  • Frame-level grouping of spectrograms, followed by clustering, is a workable alternative to patch-based encoding for capturing short intermittent sounds such as crackles.
  • The four losses (classification, clustering, cluster-similarity, and contrastive) can be jointly optimized without any single objective dominating, as the reported training curves and ablations indicate.
  • The same trained network can be converted to ONNX and run on an Android phone, so the accuracy gains translate into a deployable intelligent auscultation system.

Reading between the lines

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

  • The paper leaves implicit that its five-cluster assumption is an empirical claim: cluster assignments can be cross-tabulated against the known four class labels, and if the clusters do not track the semantic categories, the reported gain cannot be attributed to semantic clustering.
  • The paper's own noise ablation and duration statistics imply that the main error source is normal respiratory components inside annotated abnormal cycles, so tighter annotation around adventitious events would likely lift sensitivity more than further architectural changes; this is an inference from their analysis, not a result they test.
  • The group-mix contrastive mechanism is not tied to respiratory sounds; the same group-level mixing with a similarity-scaled contrastive loss could be applied to other sparse-event audio classification tasks such as heart sounds or cough detection, which the paper does not demonstrate.
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

5 major / 8 minor

Summary. The paper proposes CycleGuardian, a lightweight network for four-class respiratory sound classification on ICBHI2017, combining grouped multi-channel spectrograms (Mel, CQT, gamma), an improved deep embedding clustering module with cosine-similarity constraints on cluster features, and a group-mix contrastive learning branch, trained with a multi-loss objective. The reported headline result is Sp 82.06%, Se 44.47%, Score 63.26%, with a 38M model and no pretrained weights, claimed to be the best published no-pretrained result on the official split. The paper also presents ablation studies on the clustering and contrastive modules, an analysis of confusion between abnormal and normal classes, experiments under patient-based and stethoscope-based partitions, and an Android deployment demo.

Significance. If the reported result is stable, the paper would make a practically useful contribution: a relatively compact model that outperforms much larger transformer-based baselines without pretrained weights, together with a plausible decomposition of the gains into grouped spectrogram encoding, deep clustering, and contrastive learning. The ablation design is instructive, and the analysis in Section 4.3.1 and Section 4.4.3 identifying normal respiratory components as the main source of misclassification is a useful observation. However, the central empirical claim currently rests on a single evaluation-split-selected point estimate with no error bars, no significance testing, and no released code, and several supporting analyses do not correspond to the headline configuration. These issues are load-bearing for the claimed state-of-the-art status rather than cosmetic.

major comments (5)
  1. [Section 4.3.1, Table 2; Section 4.4.2, Table 9] The headline configuration Ours(soft cos) was selected by comparing it with Ours(cos) on the same official validation split that is then used to report the final Score. This makes 63.26% a selected maximum over at least two candidate models rather than a pre-registered single run. No multiple seeds, error bars, or significance tests are reported, so the 6.5-point advantage over ARSC-net and the 2.14-point advantage over Ours(cos) are not established as stable. The authors should report mean and standard deviation over several independent training runs, or use a proper nested validation procedure for any hyperparameter or configuration selection.
  2. [Section 4.3.1, Table 2; Section 4.4.2, Table 9] The two final configurations differ in model size (27MB vs 38MB) even though the text describes the only change as replacing cosine similarity with soft cosine similarity. This implies unstated architectural differences beyond the similarity module. The paper should specify exactly which components account for the size difference, otherwise the comparison between Ours(cos) and Ours(soft cos) is confounded.
  3. [Section 4.3.1, Fig. 7 and Table 3] The confusion matrix and per-class precision/recall/F1 report a normal-class recall of 88.92%, which matches the Sp value of Ours(cos) in Table 2, not the Sp value of the headline Ours(soft cos) model (82.06%). Tables 3 and 4 and the accompanying discussion therefore characterize a different model from the one whose results are claimed as state of the art. The per-class analysis should be recomputed for the actual headline configuration.
  4. [Section 3.2.1, Eq. (3)] The contrastive loss is a weighted similarity term rather than a normalized softmax-based contrastive objective. The paper does not explain why this objective avoids feature collapse or how the λ-weighted similarity to the source sample and to the donor sample is meant to be optimized jointly. At minimum, the authors should justify this design choice and compare it with a standard InfoNCE-style loss in the ablation study.
  5. [Section 3.2.2] The deep clustering module fixes the number of clusters at five and assumes they correspond to normal, noise, crackle, wheeze, and both, but no evidence is provided that the learned clusters align with these semantic categories. Since cluster features are fused into the global feature, a mismatch between learned clusters and intended categories could mix unrelated groups and either degrade performance or produce gains for reasons unrelated to the stated mechanism. The authors should report cluster purity or an assignment-quality measure, or provide an ablation showing how the fused clustering features behave as the number of clusters varies.
minor comments (8)
  1. [Keywords] The keyword list contains the typo "Constrative learning"; it should read "Contrastive learning."
  2. [Section 4.1] The title contains the typo "Dateset"; it should read "Dataset."
  3. [Abstract] The Abstract states the method "leads by nearly 7%," but Table 2 shows the best no-pretrained baseline is ARSC-net at 56.76%, a difference of 6.50 points; the claim should be phrased precisely.
  4. [Fig. 11] The curves showing the effect of group size on Sp and Se are plotted without error bars or repeated runs, making it difficult to tell whether the choice of 20 frames per group is robust.
  5. [Section 3.2.2, Eq. (6)] The learnable similarity matrix S is never defined or described; the authors should state its shape, initialization, and whether it is constrained (e.g., positive definite) during training.
  6. [Declarations] The Declarations section contains a paragraph about social media platform ethics and anonymizing user names that is unrelated to this study; this appears to be boilerplate text and should be removed.
  7. [Tables 8 and 9] The ablation study reports single Sp/Se/Score values without variability information; adding standard deviations would strengthen the conclusions.
  8. [Code availability] The paper states that code will be released only after acceptance; providing code or at least a detailed reproducible training configuration (random seed, augmentation schedule, optimizer settings per loss term) would substantially increase confidence in the single reported result.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found: the reported score is an empirical benchmark result, not a quantity derived from its own inputs by construction.

full rationale

The paper's central claim is an empirical classification score on the public ICBHI2017 dataset. No equation in the paper defines the result in terms of the input in a way that would make the claim true by construction. The deep clustering module (Section 3.2.2) uses a soft-assignment target pij derived from qij, and the cosine-similarity loss penalizes correlation among cluster features; these are auxiliary training objectives, not a derivation of the final Score. The contrastive loss (Eq. 3) is a self-supervised objective, and the reported Sp/Se/Score are computed by the official evaluation formula (Eq. 11) from actual predictions. There are no load-bearing self-citations: the authors' references to prior DEC and contrastive-learning work are external baselines and method components, not a chain that forces the stated 63.26% result. One legitimate experimental-design concern, noted in Section 4.4.2, is that the final 'soft cos' configuration was selected after comparing it with the 'cos' variant on the same official split used to report the headline number, and the confusion-matrix analysis in Fig. 7 and Table 3 corresponds to the rejected variant; this is an evaluation and selection-bias risk, not a circular derivation, because the reported number still comes from an actual run under the stated protocol. The unverified cluster-semantics assumption and lack of released code are correctness and reproducibility concerns, not circularity.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

This is an empirical deep learning paper with no new physical entities. The headline result depends on multiple training hyperparameters, several of which are not reported or are selected on the official validation split.

free parameters (8)
  • Group size (frames per group) = 20 frames (with 5-frame overlap)
    Selected from a sweep of 2-25 frames in Fig. 11 to maximize validation specificity/sensitivity balance; the headline result depends on this choice.
  • Clustering loss weight alpha (Eq. 10) = Not reported
    Hyperparameter balancing L_clu in the total loss; alpha also denotes the Student's t degrees of freedom in Eq. 4-5, and its value as a loss weight is never given.
  • Similarity loss weight gamma (Eq. 10) = Not reported
    Hyperparameter balancing L_cos; value not reported.
  • Contrastive temperature tau (Eq. 3) = Not reported
    Sharpness of cosine similarity in the contrastive loss; value not reported.
  • Data augmentation probabilities = Audio aug 50%, spectrogram masking ~30%
    Chosen by validation-side comparison in Table 8; these choices affect the reported score.
  • Group-mix ratio lambda distribution = Beta distribution, parameters unspecified
    Sampled per batch; distribution parameters are not stated, so the mixing schedule is underspecified.
  • Model size / capacity (38MB) = 38MB
    The lightweight architecture capacity is a design choice; it constrains performance and the paper claims it as a feature.
  • Batch size = Not reported
    A key training hyperparameter that affects contrastive learning and stability; absent from the paper.
assumptions (5)
  • domain assumption The ICBHI2017 official 6-4 split is a clean held-out evaluation; labels are correct and the official score metric is the right target.
    The entire comparison is based on this public benchmark and metric; no additional validation or error analysis is performed.
  • domain assumption Five latent clusters correspond to normal, noise, crackle, wheeze, and both.
    Section 3.2.2 fixes k=5 with semantic labels; no cluster-quality verification is provided.
  • domain assumption Grouping spectrograms into 20-frame windows preserves the temporal structure of abnormal sounds (crackles, wheezes).
    The method assumes that 20-frame groups are sufficient to capture intermittent abnormal events; no analysis of event duration vs. group size beyond Fig. 11.
  • standard math Student's t-distribution soft assignment and KL-divergence target (Eq. 4-5) are valid for clustering the embedded features.
    Imported from DEC (Xie et al. 2016); used as a standard component.
  • ad hoc to paper The soft cosine similarity with a learnable matrix S (Eq. 6) is a meaningful distance for the cluster features.
    This is a custom formulation; its behavior and regularization are not analyzed, and the matrix S is unconstrained.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CycleGuardian: A Framework for Automatic RespiratorySound classification Based on Improved Deep clustering and Contrastive Learning." pith.science (2026). https://pith.science/paper/65VHFWVL

@misc{pith2026250200734,
  author       = {Pith},
  title        = {Pith review of: CycleGuardian: A Framework for Automatic RespiratorySound classification Based on Improved Deep clustering and Contrastive Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/65VHFWVL}},
  note         = {Machine review of arXiv:2502.00734}
}
abstract

Auscultation plays a pivotal role in early respiratory and pulmonary disease diagnosis. Despite the emergence of deep learning-based methods for automatic respiratory sound classification post-Covid-19, limited datasets impede performance enhancement. Distinguishing between normal and abnormal respiratory sounds poses challenges due to the coexistence of normal respiratory components and noise components in both types. Moreover, different abnormal respiratory sounds exhibit similar anomalous features, hindering their differentiation. Besides, existing state-of-the-art models suffer from excessive parameter size, impeding deployment on resource-constrained mobile platforms. To address these issues, we design a lightweight network CycleGuardian and propose a framework based on an improved deep clustering and contrastive learning. We first generate a hybrid spectrogram for feature diversity and grouping spectrograms to facilitating intermittent abnormal sound capture.Then, CycleGuardian integrates a deep clustering module with a similarity-constrained clustering component to improve the ability to capture abnormal features and a contrastive learning module with group mixing for enhanced abnormal feature discernment. Multi-objective optimization enhances overall performance during training. In experiments we use the ICBHI2017 dataset, following the official split method and without any pre-trained weights, our method achieves Sp: 82.06 $\%$, Se: 44.47$\%$, and Score: 63.26$\%$ with a network model size of 38M, comparing to the current model, our method leads by nearly 7$\%$, achieving the current best performances. Additionally, we deploy the network on Android devices, showcasing a comprehensive intelligent respiratory sound auscultation system.

Figures

Figures reproduced from arXiv: 2502.00734 by the authors.

Figure 1
Figure 1. Flowchart of the automatic respiratory sounds classification framework. In stage 1, the obtained [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The left part shows the GFE Unit in the network, the middle shows the idea of grouping and encoding [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Group mix module: input groupi into GFE Unit to get group feature gi , and then use gi in group mix to generate new hybrid sample features for subsequent contrastive learning. controlling the concentration of the distribution. In DEC, α is typically set to a predefined value (e.g., 1), but it can also be learned. qij is termed as soft assignment, interpreting the probability of the clustering module assigning the i-… view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: The Improved Deep Embedding Clustering (IDEC) Module: including the DEC Module and Cluster [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: CycleGuardian network architecture: the red dashed box contains the main modules of the network, the [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: The upper show the Multiloss and bottom show the total loss. [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Confusion matrices corresponding to the official 6-4 divisions category than in the normal category. The F1-score for the ”both” category is also close to that of the nor￾mal category, indicating that the model performs well in detecting symptoms in the ”both” class. H…
Figure 8
Figure 8. Figure 8: Model generalization performance test, confusion matrices with heatmaps, from fold0 to fold4 patient [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Confusion matrices with heatmaps corresponding to the 8-2 divisions [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 10
Figure 10. Figure 10: Generalization test on different stethoscope devices. [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 11
Figure 11. Figure 11: Effect of the number of frames in each group [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]
Figure 12
Figure 12. Figure 12: The confusion matrix on four class without and with noise augmentation. [PITH_FULL_IMAGE:figures/full_fig_p017_12.png]
Figure 13
Figure 13. Figure 13: Respiratory sounds auscultation simulation: The left side shows the stethoscope used in this paper and [PITH_FULL_IMAGE:figures/full_fig_p018_13.png]
Figure 14
Figure 14. Figure 14: Automatic classification of recorded respiratory sounds by the model deployed on the Android [PITH_FULL_IMAGE:figures/full_fig_p019_14.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Adaptive Differential Denoising for Respiratory Sounds Classification

    eess.AS 2025-06 conditional novelty 3.0 of 10

    An Adaptive Differential Denoising network achieves a 65.53% average score on ICBHI 2017 respiratory sound classification, surpassing the previous best by 1.99%.

Reference graph

Works this paper leans on

53 extracted references · 51 canonical work pages · cited by 1 Pith paper

  1. [1]

    World health statistics 2022: monitoring health for the sdgs, sustainable development goals,

    W. H. Organization et al., “World health statistics 2022: monitoring health for the sdgs, sustainable development goals,” 2022

    work page 2022

  2. [2]

    Projections of global mor- tality and burden of disease from 2002 to 2030,

    C. D. Mathers and D. Loncar, “Projections of global mor- tality and burden of disease from 2002 to 2030,” PLoS medicine, vol. 3, no. 11, p. e442, 2006

    work page 2002

  3. [3]

    Who global progress report on tuberculosis elimination,

    E. Harding, “Who global progress report on tuberculosis elimination,” The Lancet Respiratory Medicine, vol. 8, no. 1, p. 19, 2020

    work page 2020

  4. [4]

    The global im- pact of respiratory disease–second edition. sheffield, eu- ropean respiratory society,

    F. of International Respiratory Societies, “The global im- pact of respiratory disease–second edition. sheffield, eu- ropean respiratory society,” 2017

    work page 2017

  5. [5]

    Funda- mentals of lung auscultation,

    A. Bohadana, G. Izbicki, and S. S. Kraman, “Funda- mentals of lung auscultation,” New England Journal of Medicine, vol. 370, no. 8, pp. 744–751, 2014

    work page 2014

  6. [6]

    Automatic adventitious respiratory sound analysis: A systematic review,

    R. X. A. Pramono, S. Bowyer, and E. Rodriguez-Villegas, “Automatic adventitious respiratory sound analysis: A systematic review,” PloS one, vol. 12, no. 5, p. e0177926, 2017

    work page 2017

  7. [7]

    Analysis of respiratory sounds: state of the art,

    S. Reichert, R. Gass, C. Brandt, and E. Andr` es, “Analysis of respiratory sounds: state of the art,” Clinical medicine. Circulatory, respiratory and pulmonary medicine, vol. 2, pp. CCRPM–S530, 2008

    work page 2008

  8. [8]

    New parameters for respi- ratory sound classification,

    M. Bahoura and C. Pelletier, “New parameters for respi- ratory sound classification,” in CCECE 2003-Canadian Conference on Electrical and Computer Engineering. Toward a Caring and Humane Technology (Cat. No. 03CH37436), vol. 3, pp. 1457–1460, IEEE, 2003

    work page 2003

Show all 53 references
  1. [9]

    Respirenet: A deep neural network for accurately detecting abnormal lung sounds in limited data setting,

    S. Gairola, F. Tom, N. Kwatra, and M. Jain, “Respirenet: A deep neural network for accurately detecting abnormal lung sounds in limited data setting,” in 2021 43rd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), pp. 527–530, IEEE, 2021

    2021

  2. [10]

    Lungrn+ nl: An improved ad- ventitious lung sound classification using non-local block resnet neural network with mixup data augmentation.,

    Y. Ma, X. Xu, and Y. Li, “Lungrn+ nl: An improved ad- ventitious lung sound classification using non-local block resnet neural network with mixup data augmentation.,” in Interspeech, pp. 2902–2906, 2020

    2020

  3. [11]

    Automated lung sound classification using a hy- brid cnn-lstm network and focal loss function,

    G. Petmezas, G.-A. Cheimariotis, L. Stefanopoulos, B. Rocha, R. P. Paiva, A. K. Katsaggelos, and N. Maglav- eras, “Automated lung sound classification using a hy- brid cnn-lstm network and focal loss function,” Sensors, vol. 22, no. 3, p. 1232, 2022

    2022

  4. [12]

    Adversarial fine-tuning using generated res- piratory sound to address class imbalance,

    J.-W. Kim, C. Yoon, M. Toikkanen, S. Bae, and H.- Y. Jung, “Adversarial fine-tuning using generated res- piratory sound to address class imbalance,” in Deep Generative Models for Health Workshop NeurIPS 2023, NEURAL INFORMATION PROCESSING SYSTEMS, 2023

    2023

  5. [13]

    Repaugment: Input-agnostic representation-level augmentation for respiratory sound classification,

    J.-W. Kim, M. Toikkanen, S. Bae, M. Kim, and H.-Y. Jung, “Repaugment: Input-agnostic representation-level augmentation for respiratory sound classification,” arXiv e-prints, pp. arXiv–2405, 2024

    2024

  6. [14]

    Towards enhanced clas- sification of abnormal lung sound in multi-breath: A light weight multi-label and multi-head attention classification method,

    Y.-W. Chua and Y.-C. Cheng, “Towards enhanced clas- sification of abnormal lung sound in multi-breath: A light weight multi-label and multi-head attention classification method,” arXiv e-prints, pp. arXiv–2407, 2024

    2024

  7. [15]

    Lungattn: advanced lung sound classification using attention mechanism with dual tqwt and triple stft spectrogram,

    J. Li, J. Yuan, H. Wang, S. Liu, Q. Guo, Y. Ma, Y. Li, L. Zhao, and G. Wang, “Lungattn: advanced lung sound classification using attention mechanism with dual tqwt and triple stft spectrogram,” Physiological Measurement, vol. 42, no. 10, p. 105006, 2021

    2021

  8. [16]

    A lightweight cnn model for detect- ing respiratory diseases from lung auscultation sounds using emd-cwt-based hybrid scalogram,

    S. B. Shuvo, S. N. Ali, S. I. Swapnil, T. Hasan, and M. I. H. Bhuiyan, “A lightweight cnn model for detect- ing respiratory diseases from lung auscultation sounds using emd-cwt-based hybrid scalogram,” IEEE Journal of Biomedical and Health Informatics, vol. 25, no. 7, pp. 2595...

    2020

  9. [17]

    Classification of auscultation sounds into objective spirometry findings using mvmd and 3d cnn,

    S. Gupta, M. Agrawal, and D. Deepak, “Classification of auscultation sounds into objective spirometry findings using mvmd and 3d cnn,” in 2022 National Conference on Communications (NCC), pp. 42–47, IEEE, 2022

    2022

  10. [18]

    Ofgst-swin: Swin transformer utilizing overlap fusion-based generalized s-transform for respiratory cy- cle classification,

    F. Wang, X. Yuan, J. Bao, C.-T. Lam, G. Huang, and H. Chen, “Ofgst-swin: Swin transformer utilizing overlap fusion-based generalized s-transform for respiratory cy- cle classification,” IEEE Transactions on Instrumentation and Measurement, 2024

    2024

  11. [19]

    Triple- classification of respiratory sounds using optimized s- transform and deep residual networks,

    H. Chen, X. Yuan, Z. Pei, M. Li, and J. Li, “Triple- classification of respiratory sounds using optimized s- transform and deep residual networks,” IEEE Access, vol. 7, pp. 32845–32852, 2019

    2019

  12. [20]

    Automatic classification of large-scale respira- tory sound dataset based on convolutional neural net- work,

    K. Minami, H. Lu, H. Kim, S. Mabu, Y. Hirano, and S. Kido, “Automatic classification of large-scale respira- tory sound dataset based on convolutional neural net- work,” in 2019 19th International Conference on Control, CycleGuardian 21 Automation and Systems (ICCAS), pp. 804–...

    2019

  13. [21]

    Lungbrn: A smart digital stethoscope for de- tecting respiratory disease using bi-resnet deep learn- ing algorithm,

    Y. Ma, X. Xu, Q. Yu, Y. Zhang, Y. Li, J. Zhao, and G. Wang, “Lungbrn: A smart digital stethoscope for de- tecting respiratory disease using bi-resnet deep learn- ing algorithm,” in 2019 IEEE Biomedical Circuits and Systems Conference (BioCAS), pp. 1–4, IEEE, 2019

    2019

  14. [22]

    Arsc-net: Adventitious respiratory sound classification network using parallel paths with channel-spatial at- tention,

    L. Xu, J. Cheng, J. Liu, H. Kuang, F. Wu, and J. Wang, “Arsc-net: Adventitious respiratory sound classification network using parallel paths with channel-spatial at- tention,” in 2021 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), pp. 1125–1130, IEEE, 2021

    2021

  15. [23]

    Lung sounds classification using convolutional neural net- works,

    D. Bardou, K. Zhang, and S. M. Ahmad, “Lung sounds classification using convolutional neural net- works,” Artificial intelligence in medicine, vol. 88, pp. 58– 69, 2018

    2018

  16. [24]

    Gammatone- gram based triple classification of lung sounds using deep convolutional neural network with transfer learn- ing,

    S. Gupta, M. Agrawal, and D. Deepak, “Gammatone- gram based triple classification of lung sounds using deep convolutional neural network with transfer learn- ing,” Biomedical Signal Processing and Control, vol. 70, p. 102947, 2021

    2021

  17. [25]

    Clas- sification of adventitious sounds combining cochleogram and vision transformers,

    L. D. Mang, F. D. Gonz´ alez Mart ´ ınez, D. Mar- tinez Mu˜ noz, S. Garc ´ ıa Gal´ an, and R. Cortina, “Clas- sification of adventitious sounds combining cochleogram and vision transformers,” Sensors, vol. 24, no. 2, p. 682, 2024

    2024

  18. [26]

    Feature extraction of lung sounds based on bispectrum analysis,

    S. Li and Y. Liu, “Feature extraction of lung sounds based on bispectrum analysis,” in 2010 Third International Symposium on Information Processing, pp. 393–397, IEEE, 2010

    2010

  19. [27]

    Automatic lung health screening using respiratory sounds,

    H. Mukherjee, P. Sreerama, A. Dhar, S. M. Obaidullah, K. Roy, M. Mahmud, and K. Santosh, “Automatic lung health screening using respiratory sounds,” Journal of Medical Systems, vol. 45, pp. 1–9, 2021

    2021

  20. [28]

    Convolutional neu- ral networks based efficient approach for classification of lung diseases,

    F. Demir, A. Sengur, and V. Bajaj, “Convolutional neu- ral networks based efficient approach for classification of lung diseases,” Health information science and systems, vol. 8, no. 1, p. 4, 2019

    2019

  21. [29]

    Masked modeling duo vision transformer with multi-layer feature fusion on respiratory sound clas- sification,

    B. Liu, S. Zhang, D. Takeuchi, D. Niizumi, N. Harada, and S. Makino, “Masked modeling duo vision transformer with multi-layer feature fusion on respiratory sound clas- sification,”

  22. [30]

    Patch-mix contrastive learning with audio spectrogram transformer on respi- ratory sound classification,

    S. Bae, J. Kim, W. Cho, H. Baek, S. Son, B. Lee, C. Ha, K. Tae, S. Kim, and S. Yun, “Patch-mix contrastive learning with audio spectrogram transformer on respi- ratory sound classification,” 2023

    2023

  23. [31]

    Ef- ficientvit: Lightweight multi-scale attention for high- resolution dense prediction,

    H. Cai, J. Li, M. Hu, C. Gan, and S. Han, “Ef- ficientvit: Lightweight multi-scale attention for high- resolution dense prediction,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 17302–17313, 2023

    2023

  24. [32]

    Deep learning-based lung sound analysis for intelligent stethoscope,

    D.-M. Huang, J. Huang, K. Qiao, N.-S. Zhong, H.-Z. Lu, and W.-J. Wang, “Deep learning-based lung sound analysis for intelligent stethoscope,” Military Medical Research, vol. 10, no. 1, p. 44, 2023

    2023

  25. [33]

    Patch-level contrastive embedding learning for respiratory sound classification,

    W. Song and J. Han, “Patch-level contrastive embedding learning for respiratory sound classification,” Biomedical Signal Processing and Control, vol. 80, p. 104338, 2023

    2023

  26. [34]

    Contrastive embeddind learning method for respiratory sound classification,

    W. Song, J. Han, and H. Song, “Contrastive embeddind learning method for respiratory sound classification,” in ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 1275–1279, IEEE, 2021

    2021

  27. [35]

    Pretraining respiratory sound representations using metadata and contrastive learning,

    I. Moummad and N. Farrugia, “Pretraining respiratory sound representations using metadata and contrastive learning,” in 2023 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (W ASPAA), pp. 1–5, IEEE, 2023

    2023

  28. [36]

    Stethoscope-guided supervised contrastive learn- ing for cross-domain adaptation on respiratory sound classification,

    J.-W. Kim, S. Bae, W.-Y. Cho, B. Lee, and H.-Y. Jung, “Stethoscope-guided supervised contrastive learn- ing for cross-domain adaptation on respiratory sound classification,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), ...

    2024

  29. [37]

    Unsupervised deep embedding for clustering analysis,

    J. Xie, R. Girshick, and A. Farhadi, “Unsupervised deep embedding for clustering analysis,” in International conference on machine learning, pp. 478–487, PMLR, 2016

    2016

  30. [38]

    Deep k-means: Jointly clustering with k-means and learning representa- tions,

    M. M. Fard, T. Thonet, and E. Gaussier, “Deep k-means: Jointly clustering with k-means and learning representa- tions,” Pattern Recognition Letters, vol. 138, pp. 185– 192, 2020

    2020

  31. [39]

    Joint unsupervised learning of deep representations and image clusters,

    J. Yang, D. Parikh, and D. Batra, “Joint unsupervised learning of deep representations and image clusters,” in Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 5147–5156, 2016

    2016

  32. [40]

    Deep adaptive image clustering,

    J. Chang, L. Wang, G. Meng, S. Xiang, and C. Pan, “Deep adaptive image clustering,” in Proceedings of the IEEE international conference on computer vision, pp. 5879–5887, 2017

    2017

  33. [41]

    Adaptive self-paced deep clustering with data augmentation,

    X. Guo, X. Liu, E. Zhu, X. Zhu, M. Li, X. Xu, and J. Yin, “Adaptive self-paced deep clustering with data augmentation,” IEEE Transactions on Knowledge and Data Engineering, vol. 32, no. 9, pp. 1680–1693, 2019

    2019

  34. [42]

    A frame- work for deep constrained clustering,

    H. Zhang, T. Zhan, S. Basu, and I. Davidson, “A frame- work for deep constrained clustering,” Data Mining and Knowledge Discovery, vol. 35, pp. 593–620, 2021

    2021

  35. [43]

    Agglomerative constrained clus- tering through similarity and distance recalculation,

    G. Gonz´ alez-Almagro, J. L. Suarez, J. Luengo, J.-R. Cano, and S. Garc ´ ıa, “Agglomerative constrained clus- tering through similarity and distance recalculation,” in International Conference on Hybrid Artificial Intelligence Systems, pp. 424–436, 2020

    2020

  36. [44]

    A domain transfer based data augmentation method for automated respiratory clas- sification,

    Z. Wang and Z. Wang, “A domain transfer based data augmentation method for automated respiratory clas- sification,” in ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 9017–9021, IEEE, 2022

    2022

  37. [45]

    An ensemble of deep learning frame- works for predicting respiratory anomalies.,

    L. Pham, D. Ngo, K. Tran, T. Hoang, A. Schindler, and I. McLoughlin, “An ensemble of deep learning frame- works for predicting respiratory anomalies.,” in Annual International Conference of the IEEE Engineering in Medicine and Biology Society. IEEE Engineering in Medicine and ...

    2022

  38. [46]

    Lung sound classifica- tion using co-tuning and stochastic normalization,

    T. Nguyen and F. Pernkopf, “Lung sound classifica- tion using co-tuning and stochastic normalization,”IEEE Transactions on Biomedical Engineering, vol. 69, no. 9, pp. 2872–2882, 2022

    2022

  39. [47]

    Adventitious respiratory classification using attentive residual neural networks,

    Z. Yang, S. Liu, M. Song, E. Parada-Cabaleiro, and B. W. Schuller, “Adventitious respiratory classification using attentive residual neural networks,” 2020

    2020

  40. [48]

    Cnn-moe based framework for classifi- cation of respiratory anomalies and lung disease detec- tion,

    L. Pham, H. Phan, R. Palaniappan, A. Mertins, and I. McLoughlin, “Cnn-moe based framework for classifi- cation of respiratory anomalies and lung disease detec- tion,” IEEE journal of biomedical and health informatics, vol. 25, no. 8, pp. 2938–2947, 2021

    2021

  41. [49]

    Prototype learning for interpretable respiratory sound analysis,

    Z. Ren, T. T. Nguyen, and W. Nejdl, “Prototype learning for interpretable respiratory sound analysis,” in ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 9087–9091, IEEE, 2022

    2022

  42. [50]

    Example-based explanations with adversarial attacks for respiratory sound analysis,

    Y. Chang, Z. Ren, T. T. Nguyen, W. Nejdl, and B. W. Schuller, “Example-based explanations with adversarial attacks for respiratory sound analysis,” 2022. 22 Yun et al

    2022

  43. [51]

    Deep neural network for res- piratory sound classification in wearable devices enabled by patient specific model tuning,

    J. Acharya and A. Basu, “Deep neural network for res- piratory sound classification in wearable devices enabled by patient specific model tuning,” IEEE transactions on biomedical circuits and systems, vol. 14, no. 3, pp. 535– 544, 2020

    2020

  44. [52]

    Noise masking recurrent neural network for respiratory sound classification,

    K. Kochetov, E. Putin, M. Balashov, A. Filchenkov, and A. Shalyto, “Noise masking recurrent neural network for respiratory sound classification,” in Artificial Neural Networks and Machine Learning–ICANN 2018: 27th International Conference on Artificial Neural Networks, Rhodes,...

    2018

  45. [53]

    Diagnosis of covid- 19 via acoustic analysis and artificial intelligence by monitoring breath sounds on smartphones,

    Z. Chen, M. Li, R. Wang, W. Sun, J. Liu, H. Li, T. Wang, Y. Lian, J. Zhang, and X. Wang, “Diagnosis of covid- 19 via acoustic analysis and artificial intelligence by monitoring breath sounds on smartphones,” Journal of Biomedical Informatics, vol. 130, p. 104078, 2022

    2022

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.