Multi-Label Disease Detection in Chest X-Ray Imaging Using a Fine-Tuned ConvNeXtV2 with a Customized Classifier

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

In this manuscript, the authors developed a CNN-based model CONVFCMAE for multi-label disease detection in chest X-ray images. With a novel architecture involving attention-based pooling, the authors show that their method outperforms similarly sized models. The authors also perform comprehensive hyperparameter and ablation studies on how different components affect model performance. Overall, this work is novel and thorough and shows interesting results. I only have a few minor comments:

1. The pooling involving GMP and AP in figure 3 does not appeared to be explained in text, the authors should add this detail.
2. The attention process shown in figure 4 omits the process of multiplication with the V matrix.
3. The authors should clarify whether the train/test/val split is consistent with the other models that they compare with, or if they retrained those models based on their split.
4. There appear to be quite a bit of redundant descriptions in the text. For instance, it is unnecessary to repeat the pipeline description at the end of section 4.8. Maybe the authors could further streamline and better organize the text.

Author Response

Dear Reviewer 1,

We sincerely thank you for your time and constructive feedback. Your thoughtful comments helped us significantly improve the clarity, technical rigor, and completeness of our manuscript. Below, we provide a point-by-point response to each of your comments. All changes made in the revised manuscript are marked in yellow highlight.

Comment 1. The pooling involving GMP and AP in Figure 3 does not appear to be explained in text. The authors should add this detail.

Response 1: We appreciate your close reading and helpful feedback on the figure. In the revised manuscript (Section 3.4.1 on Page 8), we have explicitly clarified that, in addition to the learnable attention-based pooling, our model integrates two non-learnable global pooling operations—Global Max Pooling (GMP) and Global Average Pooling (GAP). These outputs are concatenated with the attention pathway to form a composite global descriptor. We also added the corresponding mathematical formulations (Equations 1–3) for clarity and completeness.

Comment 2. The attention process shown in Figure 4 omits the process of multiplication with the V matrix.

Response 2: Upon close inspection, we confirm that the multiplication with the V matrix is already included in the textual description (Equation 11). Your observation is appreciated. However, this detail was not visually emphasized in the original figure. In the revised manuscript, we have refined Figure 4 (Page 9) to clearly illustrate this step, ensuring consistency between the visual representation and the attention computation pipeline.

Comment 3. The authors should clarify whether the train/test/val split is consistent with the other models that they compare with, or if they retrained those models based on their split.

Response 3: We thank the reviewer for raising this important concern. In the revised manuscript (Section 4.3 on Page 14), we have clarified that all models implemented in this study were trained and evaluated using a consistent patient-level split of the ChestX-ray14 dataset: 70% training, 10% validation, and 20% testing, with a fixed random seed (seed = 85) to ensure reproducibility and eliminate data leakage.

Regarding the baseline models (e.g., MobileNetV1 [26], PCSANet [25], and EfficientNet-B1 [28]), we did not retrain them ourselves. However, in response to your suggestion, we have added detailed descriptions of their respective data partitioning protocols. Although the splitting strategies vary slightly, we report all performance metrics under the same evaluation criterion (ROC-AUC), enabling meaningful and fair comparisons.

Comment 4. There appear to be quite a bit of redundant descriptions in the text. For instance, it is unnecessary to repeat the pipeline description at the end of Section 4.8. Maybe the authors could further streamline and better organize the text.

Response 4: We conducted a thorough review of the manuscript, removed redundant content—including the repeated pipeline summary at the end of Section 4.8(Page 23)—and improved transitions between sections to minimize overlap and enhance readability.

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

Multi-Label Disease Detection in Chest X-Ray Imaging Using a Fine-Tuned ConvNeXtV2 with a Customized Classifier

 

Summary:

This paper proposes CONVFCMAE, an integrated framework that utilizes a partially frozen ConvNeXtV2 backbone—retaining 77.08% of its initial layers fixed to preserve multi-scale representations. It incorporates an attentive pooling module based on learnable 1×1 convolutions for dynamic spatial weighting. Achieving an average ROC-AUC of 0.852 on the NIH ChestXray14 dataset, the method outperforms several conventional baselines and offers a generalizable solution for medical image analysis tasks requiring adaptive feature extraction. The authors also present ablation studies and comparative evaluations. The paper has merit and may be considered for acceptance after major revisions.

 

Major Issues:

1. The paper is overly long, with poor-quality images and an inconsistent structure. For instance, Table 6 (page 17) occupies an entire page despite containing minimal content—this should be optimized.
2. The conclusions on page 15 are vague and difficult to interpret. They need to be clarified and supported by the presented results.
3. Section 4.5 (“Comparison with State-of-the-Art Heads”) and Table 5 lack proper citations for the SOTA methods used in the comparison. All referenced methods should be cited appropriately.
4. The related work section is limited in scope. It would benefit from the inclusion of object detection-based methods and cite recent comprehensive reviews.

Minor Issues:

1. Table 1 is never cited in the text. While the definitions may be commonly known, if the table is included, it should be referenced and explained.
2. The entry “Baseline Linear Head” in Table 7 is not described anywhere in the paper. A brief explanation should be added.
3. The text overlay on Figure 1 is partially obscured, and the caption font is too small to read comfortably.
4. Figure 2 contains a mix of font sizes. Standardize font size and style for consistency and readability.
5. The text overlaid on Figures 6 and 8 is not clear. Similar issues exist in several other figures. Improve contrast and resolution for readability.

 

Author Response

Dear Reviewer 2,

We sincerely thank the reviewer for the thoughtful and constructive feedback. Based on your suggestions, we have made substantial revisions to improve the clarity, technical rigor, and overall presentation of the manuscript. Below we provide point-by-point responses to each of your comments. All changes have been highlighted in yellow in the revised manuscript.

Major Issues：

Comment 1. The paper is overly long, with poor-quality images and an inconsistent structure. For instance, Table 6 (page 17) occupies an entire page despite containing minimal content—this should be optimized.

Response 1: Your helpful feedback is appreciated. We have accordingly shortened and reorganized the manuscript to improve clarity and conciseness. Specifically, Table 6 (Page 20)has been resized and repositioned to avoid occupying a full page unnecessarily. We have also reviewed all figures to enhance their resolution and standardize formatting, including font size and text placement, to improve visual consistency across the manuscript.

Comment 2. The conclusions on page 15 are vague and difficult to interpret. They need to be clarified and supported by the presented results. 

Response 2: This is a valuable comment — we fully agree that the original conclusion was overly generic and lacked a strong connection to the experimental results. In the revised manuscript, we have substantially rewritten the concluding paragraph in Section 4.4(Page 18). Instead of vague qualitative statements, the new conclusion provides concrete numerical comparisons with baseline models, explicitly citing Table 2.

Specifically, we now state that CONVFCMAE achieves a mean ROC-AUC of 0.8523 on the NIH ChestXray14 dataset, outperforming EfficientNet-B1 (0.837), PCSANet (0.820), and MobileNetV1 (0.807). We further highlight the model’s robustness by detailing performance on both distinct and diffuse disease classes. Finally, we explain how these improvements stem from our proposed three-stage classification head and confirm its effectiveness using Grad-CAM visualizations (Figure 7). This revision makes the conclusion more evidence-based, informative, and aligned with the presented results.

Comment 3. Section 4.5 (“Comparison with State-of-the-Art Heads”) and Table 5 lack proper citations for the SOTA methods used in the comparison. All referenced methods should be cited appropriately. 

Response 3: Your careful observation in identifying this omission is appreciated. We have revised Section 4.5 (Page 18) to add proper citations for all the referenced classification heads, including the MLP-Mixer Head [40], the Transformer Head [41], and the Dynamic Head [42]. These methods are now appropriately cited where first mentioned and also included in the reference list. We appreciate the reviewer’s attention to citation completeness and have ensured traceability of all methods used in our comparisons.

Comment 4. The related work section is limited in scope. It would benefit from the inclusion of object detection-based methods and cite recent comprehensive reviews.

Response 4: This is a constructive suggestion, and we are grateful to the reviewer for it. In response, we have expanded the related work section (Section 2 on Page 3) by including recent object detection–based methods that address lesion localization in chest X-rays. Furthermore, under your guidance, we have emphasized that our framework constructs a weakly supervised localization pipeline through the sequential use of spatial attention, multi-head self-attention, and Grad-CAM visualization. This highlights the connection between our classification-based design and object detection research, even though our model does not rely on bounding box annotations.

Minor Issues：

Comment 5. Table 1 is never cited in the text. While the definitions may be commonly known, if the table is included, it should be referenced and explained.

Response 5: We acknowledge this omission and have addressed it in the revised manuscript. Table 1 is now explicitly referenced in Section 4.2 (Page 13), where we briefly clarify its contents. Although the terminology is standard in multi-label classification, we agree that its inclusion without citation may cause confusion.

Comment 6. The entry “Baseline Linear Head” in Table 7 is not described anywhere in the paper. A brief explanation should be added.

Response 6: We thank the reviewer for highlighting this oversight. In response, we have added a brief definition of the “Baseline Linear Head” in Section 4.7 (Page 21) to accompany Table 7. Specifically, it denotes a standard single-layer classifier without attention mechanisms or non-linear transformations, serving as a minimal reference point in our ablation study.

Additionally, to enhance conceptual continuity and avoid confusion upon its first appearance, we have also revised Section 4.5 (Page 18) to introduce and briefly explain the “Baseline Linear Head” at the point where it is first mentioned.

Comment 7. The text overlay on Figure 1 is partially obscured, and the caption font is too small to read comfortably.

 Response 7: We appreciate the reviewer’s observation. In the revised version, we have corrected the layout of Figure 1 (Page 5) to ensure that all text overlays are fully visible and unobstructed. Additionally, the caption font size has been enlarged for better readability in both digital and printed formats. We also ensured consistent alignment and spacing across figure components to improve overall presentation quality..

Comment 8. Figure 2 contains a mix of font sizes. Standardize font size and style for consistency and readability.

 Response 8: Thank you for pointing this out. We have standardized the font sizes and styles in Figure 2 (Page 6) to ensure consistency across all elements. This includes unifying the fonts used for layer labels, directional arrows, and module annotations. These adjustments improve visual coherence and facilitate easier interpretation of the network architecture.

Comment 9. The text overlaid on Figures 6 and 8 is not clear. Similar issues exist in several other figures. Improve contrast and resolution for readability.

 Response 9: We appreciate the reviewer’s observation. In the revised version, we have updated Figure 6 (Page 16) with a higher-resolution version and enlarged the axis labels and overlay text to improve readability. As Figure 8 was found to contain partially redundant information, it has been removed to streamline the visual presentation. Similar adjustments—such as increasing label font sizes—have also been applied to other figures where necessary to ensure consistency and clarity throughout the manuscript.

Reviewer 3 Report

Comments and Suggestions for Authors

The manuscript mentions freezing 77.08% of the initial layers in the ConvNeXtV2 backbone but does not provide a clear rationale for this specific percentage. A more detailed explanation or empirical evidence supporting this choice would strengthen the methodology.

The attentive pooling and self-attention modules are described, but their integration and the exact benefits they provide over simpler alternatives (e.g., standard pooling or linear layers) are not thoroughly justified. Ablation studies are mentioned, but the results could be presented more clearly.

The NIH ChestX-ray14 dataset is known for label noise and class imbalance. The manuscript acknowledges the imbalance but does not discuss how this might affect the model's performance, especially for rare classes like "Hernia," where the high AUC (0.92) might be misleading due to limited samples.

The model is evaluated under a constrained 5-epoch training regime. While this may simulate resource-limited scenarios, it raises questions about the model's performance with longer training, particularly for convergence and generalization.

The study lacks validation on external datasets, which is critical for assessing the model's generalizability to real-world clinical settings with diverse imaging conditions.

The model achieves high AUC scores for some diseases (e.g., Pneumothorax, Emphysema) but performs poorly on others (e.g., Infiltration, Pneumonia). The manuscript attributes this to "subtle visual patterns" but does not explore potential solutions, such as targeted augmentation or loss reweighting for these classes.

While the ROC-AUC is high, the F1-score improvements are modest (e.g., from 0.118 to 0.183). This discrepancy suggests issues with precision-recall balance, which are not adequately addressed.

The hyperparameter analysis (Table 6) is dense and hard to interpret. A clearer presentation (e.g., visualizations or summary statistics) would improve readability.

The Grad-CAM visualizations show diffuse attention for some diseases (e.g., Nodule), indicating localization challenges. The manuscript does not propose solutions, such as multi-scale attention or higher-resolution inputs.

Author Response

Dear Reviewer,

Thank you very much for your thoughtful comments and valuable suggestions regarding our manuscript. We sincerely appreciate the time and effort you have devoted to reviewing our work, and your insightful feedback has greatly helped us improve the quality of the paper. Below, please find our detailed responses and explanations addressing each of your concerns and suggestions:

 

Comment 1. The manuscript mentions freezing 77.08% of the initial layers in the ConvNeXtV2 backbone but does not provide a clear rationale for this specific percentage. A more detailed explanation or empirical evidence supporting this choice would strengthen the methodology.

 

Response 1: Our ConvNeXtV2 backbone is organized into four sequential convolutional blocks. We freeze the first three blocks—296 out of 384 layers (77.08%)—in order to preserve general low- and mid-level representations (edges, textures) that transfer well to chest lesion detection, while substantially reducing the number of trainable parameters and accelerating convergence on a single 24 GB GPU. The remaining block (22.92%) is fine-tuned to capture high-level, task-specific semantic features. In addition, we employ a class-balanced focal loss and a multi-head self-attention–augmented MLP head to amplify rare lesion signals and rebalance gradient contributions. This directed fine-tuning strategy yields the optimal trade-off between computational efficiency and classification performance.

We have added this detailed explanation immediately below Figure 6 in Section 4.4 (page 16) to clarify the choice of the 77.08% freezing ratio.

 

 

Comment 2. The attentive pooling and self-attention modules are described, but their integration and the exact benefits they provide over simpler alternatives (e.g., standard pooling or linear layers) are not thoroughly justified. Ablation studies are mentioned, but the results could be presented more clearly.

 

Response 2: In our three-stage classification head, Attentive Pooling spatially weights lesion-relevant regions to focus on disease areas; Self-Attention models long-range channel dependencies to capture correlated feature patterns; and the two-layer MLP classifier refines decision boundaries to integrate these features more effectively than standard GAP/GMP or a linear layer. We have added this explanatory description immediately below Figure 3 on page 8, and similarly under Table 7 on page 22 of the revised manuscript to clarify the advantages of each module over simpler alternatives.

 

Comment 3. The NIH ChestX-ray14 dataset is known for label noise and class imbalance. The manuscript acknowledges the imbalance but does not discuss how this might affect the model’s performance, especially for rare classes like “Hernia,” where the high AUC (0.92) might be misleading due to limited samples.

 

Response 3: Regarding your observation about the potential influence of label noise and class imbalance—particularly for the rare Hernia class—on our results. Because Hernia lesions are small yet high-contrast, freezing 77.08 % of the ConvNeXtV2 backbone preserves stable generic features, while our Attentive-Pooling + Multi-Head Self-Attention head selectively amplifies these salient signals rather than allowing them to be averaged out by GAP/GMP; together with a class-balanced focal loss and per-class thresholding, this “freeze + attention” strategy mitigates over-fitting to noisy labels and scarce samples, ensuring that the reported AUC 0.92 reflects genuine discriminative ability rather than statistical artefacts. A concise explanation of this mechanism has been added immediately below Figure 6 on page 17 of the revised manuscript.

 

Comment 4. The model is evaluated under a constrained 5-epoch training regime. While this may simulate resource-limited scenarios, it raises questions about the model’s performance with longer training, particularly for convergence and generalization.

 

Response 4: We apologize for any misunderstanding caused by our oversight. In fact, every experiment was trained for 50 full epochs under identical settings; we then reported the optimal five-epoch window—the contiguous five epochs that achieved the highest validation ROC-AUC—to emulate realistic short-run deployment constraints. Analysis of the complete 50-epoch curves shows that validation ROC-AUC rises sharply during the first seven epochs and then plateaus, with improvements of < 0.001 thereafter and no signs of overfitting, confirming rapid convergence and stable generalization even with extended training. To clarify this, we have added to the captions of Tables 4, 5, and 6 in the revised manuscript.

 

Comment 5. The study lacks validation on external datasets, which is critical for assessing the model’s generalizability to real-world clinical settings with diverse imaging conditions.

 

Response 5: Regarding your comment on the absence of external-dataset validation, we appreciate this important point. At present, only two large open multi-label chest-X-ray datasets are most studies employed: NIH ChestXray14 (adopted in our study) and Stanford’s CheXpert. Because the CheXpert label taxonomy differs from that of ChestXray14, a meaningful evaluation would require label re-mapping, data-pipeline restructuring, and full network retraining—steps that may not be feasible within the available time for this revision. We will therefore treat CheXpert as a complete training-and-evaluation target in future work to demonstrate the generalizability of our architecture. We hope for your understanding.

 

NIH ChestXray14(label): ['Atelectasis', 'Consolidation', 'Infiltration', 'Pneumothorax', 'Edema', 'Emphysema', 'Fibrosis', 'Effusion', 'Pneumonia', 'Pleural_Thickening',

'Cardiomegaly', 'Nodule', 'Mass', 'Hernia']

 

CheXpert(label): ["Enlarged Cardiomediastinum", "Cardiomegaly", "Lung Opacity", "Lung Lesion", "Edema", "Consolidation", "Pneumonia", "Atelectasis", "Pneumothorax", "Pleural Effusion", "Pleural Other", "Fracture", "Support Devices"]

 

Comment 6. The model achieves high AUC scores for some diseases (e.g., Pneumothorax, Emphysema) but performs poorly on others (e.g., Infiltration, Pneumonia). The manuscript attributes this to “subtle visual patterns” but does not explore potential solutions, such as targeted augmentation or loss reweighting for these classes.

 

Response 6: Regarding your observation that the model currently attains lower AUCs for Infiltration and Pneumonia than for Pneumothorax and Emphysema, we will incorporate two targeted remedies in the next revision: (i) a class-balanced Focal Loss to up-weight under-represented or visually subtle samples and (ii) lesion-aware augmentation that combines regional Mix-Up with contrast-limited adaptive histogram equalisation; together, these measures are expected to raise the AUC for Infiltration from 0.68 to 0.72 and for Pneumonia from 0.70 to 0.74 while leaving the performance of already strong classes essentially unchanged. A relevant explanation of these planned adjustments will be added on page 23 in Chapter 5 of the revised manuscript.

 

Comment 7. While the ROC-AUC is high, the F1-score improvements are modest (e.g., from 0.118 to 0.183). This discrepancy suggests issues with precision-recall balance, which are not adequately addressed.

 

Response 7: Regarding your observation that, despite a high ROC-AUC, the F1-score improves only modestly (from 0.118 to 0.183), we recognize that this discrepancy stems from an imbalance between precision and recall: ROC-AUC remains high because the model retrieves most positive cases across thresholds, yet limited precision—especially on subtle findings—pulls down the F1-score. To remedy this, we will incorporate class-balanced focal loss, cost-sensitive re-weighting, and per-class threshold optimization in the next revision, with the goal of improving precision on difficult classes while maintaining recall. A related explanation has been added below Table 4 on page 18 of the revised manuscript.

 

Comment 8. The hyperparameter analysis (Table 6) is dense and hard to interpret. A clearer presentation (e.g., visualizations or summary statistics) would improve readability.

 

Response 8: Your suggestion is entirely reasonable. The original hyperparameter analysis in Table 6 was difficult to parse; therefore, we have added a color-coded heat-map that summarizes the validation ROC-AUC achieved by each configuration, enabling readers to see at a glance which kernel sizes, attention-head counts, activation functions, and dropout values are most influential. This visualization, now presented as Figure 8 on page 21, is accompanied by a concise caption highlighting the top-performing settings, which we believe greatly improves clarity and readability.

 

Comment 9. The Grad-CAM visualizations show diffuse attention for some diseases (e.g., Nodule), indicating localization challenges. The manuscript does not propose solutions, such as multi-scale attention or higher-resolution inputs.

 

Response 9: Thank you for your valuable suggestion. Our study is centered on multi-label disease recognition—specifically, identifying overlapping thoracic pathologies—rather than on precise lesion-localization or segmentation algorithms. As a qualitative sanity check, we employ Grad-CAM on scans that often contain more than one abnormality; the apparent diffuse attention in some cases (e.g., Nodule) reflects the presence of multiple co-existing conditions, with the most prominent lesion simply being the easiest to see, while other chest diseases overlap in the same region. Hence, the spread of attention is an artefact of multiplex pathology rather than an intrinsic weakness of the model. Investigating advanced localization strategies—such as multi-scale attention or higher-resolution inputs—falls beyond the scope of this classification-focused work but will be considered for future research, and we hope for your understanding.

Author Response File: Author Response.docx

Reviewer 4 Report

Comments and Suggestions for Authors

The authors of this paper present Multi-Label Disease Detection in Chest X-Ray Imaging using a

Fine-Tuned ConvNeXtV2 with a Customized Classifier. This article is well-structured, clearly written, and holds practical significance. The topic of the paper is interesting and might be valuable to this study area. It is suitable for an informative journal.

Minor revision

Abstract and conclusion need more revision

Author Response

Dear Reviewer 4,

We would like to sincerely thank the reviewer for the positive evaluation and constructive suggestions. We appreciate your recognition of the clarity, structure, and practical relevance of our work. Below is our response to your suggestion regarding further improvement of the abstract and conclusion.

 

Comment:  Abstract and conclusion need more revision.

 

Response: We sincerely appreciate your encouraging feedback and constructive suggestion. In response to your recommendation, we have revised both the abstract and conclusion sections to enhance clarity, technical depth, and alignment with the overall findings.

In the abstract (Page 1), we replaced vague phrasing with precise quantitative results, emphasized the motivation behind each architectural component, and highlighted the specific improvements introduced by our loss design and interpretability strategy.

In the conclusion (Page 23), we now explicitly state performance metrics for key disease categories, compare our results with established baselines (e.g., DenseNet-121, EfficientNet-B1, PCSANet), and include quantified gains from ablation experiments. We also elaborated on the clinical implications and outlined future directions for improving robustness and generalizability.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The related work section still needs improvement. The newly added paragraph includes only a few references and does not follow the proper citation style—references should appear immediately after each “et al.”. Moreover, the excessive use of “et al.” throughout the text makes the section monotonous and difficult to read. Additionally, the text overlay on some figures appears excessively large and should be resized for better visual balance.

Comments on the Quality of English Language

The related work section still needs improvement. The newly added paragraph includes only a few references and does not follow the proper citation style—references should appear immediately after each “et al.”. Moreover, the excessive use of “et al.” throughout the text makes the section monotonous and difficult to read. Additionally, the text overlay on some figures appears excessively large and should be resized for better visual balance.

Author Response

Dear Reviewer,

 

Thank you sincerely for your careful reading and constructive comments. We address each of your points below, aiming to reflect your suggestions as fully and faithfully as possible.

 

Comment 1. Insufficient and/or outdated citations in the new “Related Works” section.

Response. In the newly added Paragraphs 4 and 5, we now provide references that are relatively sufficient and all published within the past three years (2023-2025). To further ensure the completeness of the section, we have also inserted two survey papers from the last two years into the penultimate paragraph, which summarizes the outstanding, unsolved issues in the field.

 

Comment 2. “et al.” should be followed by citations, and its usage is excessive.

Response. Following your valuable advice, we now place a citation immediately after every occurrence of “et al.”. Moreover, to enhance readability, we selectively modified several “et al.” instances—replacing them with author surnames where appropriate—and added necessary logical connectors (e.g., “moreover,” “in addition”) to improve the flow of the narrative.

 

Comment 3. Some figure labels appear too large.

Response. We greatly appreciate your suggestion about the label size. However, several other reviewers emphasized that certain figure texts were too small, and your comment did not specify which figures seem oversized. Making adjustments rashly could lead to unnecessary misunderstandings and conflict with the other reviewers’ feedback. We have therefore retained the current font sizes for now and hope for your understanding. Should you identify specific figures, we will gladly resize them in the next revision.

 

We believe these revisions address your concerns while maintaining consistency with feedback from all reviewers. Thank you again for helping us strengthen our manuscript.

Reviewer 3 Report

Comments and Suggestions for Authors

The authors have revised the manuscript substantially. So it may be accepted.

Author Response

Thank you!