Pith. sign in

REVIEW 4 major objections 6 minor 19 references

Comprehensive Lung Disease Detection Using Deep Learning Models and Hybrid Chest X-ray Data with Explainable AI

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that merging four chest X-ray datasets into a hybrid set lets standard pretrained CNNs reach 99% accuracy in classifying COVID-19, pneumonia, lung opacity, and normal cases, with LIME identifying misclassification causes.

desk verdict The hybrid-dataset idea is legitimate, but the paper never rules out cross-dataset image leakage, so the headline 99% accuracy does not establish generalization. read the letter →

arxiv 2505.16028 v2 pith:6UPSOT2I submitted 2025-05-21 eess.IV cs.CV

Shuvashis Sarker , Shamim Rahim Refat , Faika Fairuj Preotee , Tanvir Rouf Shawon , Raihan Tanvir This is my paper
classification eess.IVcs.CV
keywords lungdiseasedetectionchestX-rayhybriddatasettransferlearningVGG16DenseNet121LIMEexplainableAI
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 tries to establish that merging four chest X-ray datasets—two collected in Bangladesh and two drawn from global sources—into a single hybrid dataset improves how well deep learning models detect COVID-19, pneumonia, lung opacity, and normal lung conditions. On the hybrid dataset, VGG16, Xception, ResNet50V2, and DenseNet121 each reach 99% test accuracy, which the authors interpret as strong generalization across diverse populations. The paper further claims that LIME explanations of misclassified images reveal which image regions mislead the model, adding interpretability to otherwise black-box predictions. A sympathetic reader would care because the result suggests that pooling regional and global imaging data, rather than designing new architectures, could be a practical route to broadly usable and interpretable lung-disease screening tools.

What carries the argument

The load-bearing object is the proposed hybrid dataset, built by merging Dataset 1 (a balanced augmented COVID-19 chest X-ray collection), Dataset 2 and Dataset 3 (Bangladeshi hospital chest X-ray collections), and Dataset 4 (an aggregation of 15 public datasets with viral and bacterial pneumonia merged into a single pneumonia class). Class names are normalized across sources, images are resized and split 80:10:10, and off-the-shelf pretrained CNNs (VGG16, VGG19, InceptionV3, Xception, ResNet50V2, InceptionResNetV2, MobileNetV2, DenseNet121) are fine-tuned with custom dense heads. LIME then explains individual predictions by locally approximating the black-box model with an interpretable linear model over superpixels.

What would settle it

Compute perceptual hashes or run near-duplicate detection across the hybrid training and test splits; if overlapping or augmented copies of the same source images appear in both, the reported 99% accuracy would reflect memorization rather than generalization. Alternatively, retrain the same models on the hybrid dataset with Dataset 1 and Dataset 4 removed one at a time; a large drop in test accuracy would indicate the result depends on data leakage from those aggregate sources.

Watch

Extended reading notes

Core claim

The central claim is that a hybrid dataset formed by combining four publicly available chest X-ray collections yields near-ceiling performance from standard pretrained convolutional networks. On the hybrid test split, VGG16, Xception, ResNet50V2, and DenseNet121 each achieve 99% accuracy, with precision, recall, and F1-score at 0.99, matching or exceeding their performance on any single source dataset. The authors attribute this to the diversity of the pooled data, which lets models learn features that transfer across regional populations. They also use LIME on selected misclassified images and report that the highlighted green regions correspond to the class the model wrongly favored, providing a partial window into why errors occur.

Load-bearing premise

The accuracy claims rest on the assumption that the 10% hybrid test split contains no near-duplicate or overlapping images from the four source datasets, since no de-duplication is described, and Dataset 1 is an augmented version of the COVID-19 Radiography Database while Dataset 4 aggregates 15 public datasets, so the same patient or source image could plausibly appear in both training and test splits.

Editorial extensions

If this is right

  • If the hybrid-data claim holds, pooling diverse regional and global chest X-ray repositories could let standard pretrained models approach the practical accuracy ceiling without custom architectures.
  • The 99% accuracy on the hybrid test split implies the four models generalize across Bangladeshi and global imaging conditions, supporting deployment of the same model in multiple geographies.
  • LIME's ability to highlight the regions driving misclassifications gives radiologists a concrete artifact to audit and potentially correct model errors.
  • The method's simplicity means it can be reproduced quickly on other public imaging repositories, offering a low-cost template for building interpretable screening tools.

Reading between the lines

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

  • The paper does not demonstrate generalization beyond the pooled test split; a stronger test would be training on the hybrid set and evaluating on an entirely unseen external dataset, which the authors list as future work.
  • Because Dataset 4 already aggregates 15 public datasets and Dataset 1 is an augmented version of a widely used database, the hybrid set's diversity may be overstated without an overlap analysis; the reported accuracy could partly reflect duplicated images across splits.
  • LIME is applied to only five misclassified examples, so the interpretability claim is anecdotal rather than a systematic evaluation of where and why the model fails.
  • The result that four different pretrained models all land at exactly 99% accuracy on the hybrid test split is uniform enough to warrant checking whether the test split is too easy or too similar to training images.
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

4 major / 6 minor

Summary. The manuscript proposes a hybrid chest X-ray dataset created by merging four public datasets (two from Bangladesh and two global), evaluates nine deep-learning and transfer-learning models on each individual dataset and on the hybrid, and uses LIME to visualize misclassifications. The central claim is that the hybrid dataset 'significantly enhances model accuracy and generalizability,' with VGG16, Xception, ResNet50V2, and DenseNet121 reaching 99% accuracy on the hybrid test split. The paper also compares its results with prior work and closes with limitations and future directions.

Significance. The question whether pooling regionally diverse chest X-ray data improves generalization is practically important, and the paper covers a reasonable range of standard architectures. However, the evidence for the central claim is not currently persuasive: the hybrid test split may contain near-duplicates of training images because Dataset 1 is an augmented version of the COVID-19 Radiography Database while Dataset 4 aggregates 15 public datasets with no de-duplication described; the reported metrics are selected as the best of five runs with no confidence intervals or significance tests; and preprocessing differs across datasets (128x128 vs 256x256). No code or data artifacts are provided, so the 99% numbers cannot be independently reproduced. If the authors fix these issues with a de-duplicated, statistically transparent evaluation, the paper could become a useful benchmark, but the current manuscript does not support its headline claim.

major comments (4)
  1. [§3.1–3.2, Table 1] The claimed 99% hybrid accuracy is not trustworthy without demonstrating that the hybrid split is free of near-duplicate images. Dataset 1 is described as an augmented version of the COVID-19 Radiography Database, and Dataset 4 is described as an aggregation of 15 public datasets, but those 15 sources are never listed and no overlap check is performed. Since the 80:10:10 split is applied after concatenation, the same original chest X-ray (or an augmented copy) can appear in the training portion contributed by one dataset and the test portion contributed by another. Given that Dataset 1 supplies about 30k of the roughly 42k hybrid images, models could memorize source-specific artifacts and score near ceiling on leaked test images. The authors must list the constituent sources of Dataset 4, perform a de-duplication step (e.g., perceptual hashing) across all four datasets, and report hybrid results on the de-duplicated split.
  2. [§4, Table 1] The phrase 'significantly enhances' in the abstract and Section 4 is not backed by any statistical analysis. Section 4 states that each model 'runs five times, and the best outcomes are recorded,' which selects the most favorable seed and inflates accuracy; no standard deviations, confidence intervals, or hypothesis tests are reported. The claim of significance should be supported by reporting all five runs (or mean and standard deviation) and by performing a paired test (e.g., McNemar's test) between the individual-dataset and hybrid-dataset classifications. In addition, Dataset 1 already reaches 99% for most models, so a ceiling effect makes it impossible to conclude from Table 1 alone that the hybrid dataset is responsible for the improvement.
  3. [§3.2] The preprocessing pipeline contains a confound. According to Section 3.2, 'the images from Dataset 1 and the Hybrid Dataset were resized to 128x128 pixels, while the images from Datasets 2, 3, and 4 were standardized to 256x256 pixels.' Since the hybrid dataset contains images from all four sources, it is unclear whether images from Datasets 2/3/4 are resized to 256 and then to 128, or are only resized to 128. Either way, the comparison between individual-dataset results and hybrid results is not resolution-controlled. The authors should specify a single preprocessing pipeline and ideally repeat the key comparisons at both resolutions to show that the conclusions are not an artifact of resizing.
  4. [§3.1, Table 1] The hybrid dataset is dominated by Dataset 1 (29,998 of about 42,173 images), so the near-perfect hybrid accuracy in Table 1 may simply reflect the distribution of the largest source rather than improved cross-region generalization. The authors should report test accuracy broken down by source dataset within the hybrid split, and consider whether a balanced hybrid (or a per-source stratified evaluation) changes the conclusions.
minor comments (6)
  1. [Abstract and Introduction] The writing contains numerous grammatical and typographical errors (for example, 'Palmo aka lungs' and 'The ability to apply of findings'); the manuscript needs careful copyediting.
  2. [§4] Section 4 says the models are implemented using 'Keras 3' and 'TensorFlow 4'; TensorFlow 4 does not exist, and the version should be corrected.
  3. [§5.1, Figure 3] The LIME analysis is purely anecdotal; a quantitative evaluation of explanation quality (e.g., localization against expert annotations, or agreement metrics on a larger set of misclassified examples) would be needed to support the interpretability claims.
  4. [Table 2] Table 2 compares the proposed method with prior work at different accuracies but on different datasets and class splits; this should be framed only as a coarse reference, not as a direct comparison.
  5. [General] No data availability or code availability statement is provided, which limits reproducibility given the reliance on specific public datasets.
  6. [References] Reference [19] appears to include one of the present authors (T. R. Shawon) but is cited as ordinary prior work; the self-citation should be flagged or removed according to journal policy.

Circularity Check

1 steps flagged · score 2.0 of 10

No significant circularity: the accuracy/generalizability claims are empirical benchmark results on held-out splits; the only self-reference is a routine, non-load-bearing citation.

  1. other [Section 5.1 (Explainable AI / LIME) and References [19]]
    "LIME functions by executing four fundamental procedures: permuting input data to generate artificial data points, predicting classes for these points, calculating their similarity to the original data, and employing a linear classifier to identify the most significant features that influence the model's decisions. This method enhances the transparency and comprehensibility of AI models [19]."

    This is a minor self-citation: reference [19] is authored by Maliyat, Sultana, Sultana, Saha Joy, and Shawon, where Shawon is one of the present co-authors. It is cited only to support a general statement that LIME enhances transparency, and it plays no role in deriving the reported 99% accuracies or the hybrid-dataset generalizability claim. Because the central claims rest on train/validation/test splits rather than on this citation, the self-citation is not load-bearing; per the rubric it warrants a score of 2 rather than 0, but it is not circular.

full rationale

The paper's central claims are empirical rather than derivational. Models are trained on an 80% split of each dataset and evaluated on a 10% held-out test split (Section 3.2), and the hybrid-dataset performance is measured on a similarly held-out hybrid test split. No equation in the paper defines the reported accuracy in terms of the inputs: no fitted parameter is renamed as a prediction, and no uniqueness theorem is imported to force a choice. The only quasi-circularity-adjacent item is the routine self-citation in Section 5.1, which is not load-bearing. Concerns about possible image overlap between Dataset 1 and the aggregated Dataset 4 (and hence a contaminated hybrid test split) are a data-leakage/correctness risk, not a circularity of the derivation: the models are still evaluated on images not used in that training run, and the claimed generalization would be invalidated only if duplicates exist, which is an empirical data-quality question the paper should answer but not a logical reduction of the claim to its own inputs. Similarly, the statement that each model runs five times and the best outcome is recorded (Section 4) is run selection, not fitting a parameter and calling it prediction. Accordingly, the circularity score is low.

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

The central empirical claim rests on pooling assumptions and protocol choices rather than derived constants. The most consequential are the unverified absence of train/test leakage and the best-of-five reporting, both of which can materially inflate accuracy.

free parameters (3)
  • Best-of-five run selection = Best accuracy over five runs per model and dataset
    Used throughout Table 1; selecting the best run inflates reported accuracy and hides variance, so the central claim depends on this protocol.
  • Fully connected layer sizes and dropout rates = 1024 and 512 neurons; dropout 0.5 and 0.25
    Hand-chosen for all pretrained models in Section 3.3 with no sensitivity analysis, and they affect the reported accuracy.
  • Per-dataset image resize targets = 128x128 for Dataset 1 and hybrid; 256x256 for Datasets 2, 3, 4
    Section 3.2 defines inconsistent resizing; this hand-chosen preprocessing choice can affect which patterns are preserved and biases the hybrid comparison because the hybrid resolution matches only Dataset 1.
assumptions (4)
  • domain assumption Naive pooling of the four datasets is a valid training distribution, and label definitions are consistent after renaming Viral Pneumonia to Pneumonia.
    Section 3.1 merges datasets and changes class names but does not verify label quality or harmonize annotation protocols across sources.
  • domain assumption No near-duplicate or overlapping images exist between the training and test splits of the hybrid dataset.
    Assumed in the 80:10:10 split in Section 3.2; overlap is plausible because Dataset 1 augments the COVID-19 Radiography Database and Dataset 4 aggregates 15 public datasets.
  • ad hoc to paper Best-of-five run accuracy is a stable estimator of model performance.
    Section 4 says each model runs five times and the best outcomes are recorded, so the reported numbers are upper envelopes rather than typical performance.
  • ad hoc to paper Resizing images to different resolutions across datasets does not bias the comparison.
    Dataset 1 and hybrid images are 128x128 while Datasets 2, 3, 4 are 256x256, so the hybrid resolution matches only one source and may systematically affect which datasets contribute to performance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Comprehensive Lung Disease Detection Using Deep Learning Models and Hybrid Chest X-ray Data with Explainable AI." pith.science (2026). https://pith.science/paper/6UPSOT2I

@misc{pith2026250516028,
  author       = {Pith},
  title        = {Pith review of: Comprehensive Lung Disease Detection Using Deep Learning Models and Hybrid Chest X-ray Data with Explainable AI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6UPSOT2I}},
  note         = {Machine review of arXiv:2505.16028}
}
read the original abstract

Advanced diagnostic instruments are crucial for the accurate detection and treatment of lung diseases, which affect millions of individuals globally. This study examines the effectiveness of deep learning and transfer learning models using a hybrid dataset, created by merging four individual datasets from Bangladesh and global sources. The hybrid dataset significantly enhances model accuracy and generalizability, particularly in detecting COVID-19, pneumonia, lung opacity, and normal lung conditions from chest X-ray images. A range of models, including CNN, VGG16, VGG19, InceptionV3, Xception, ResNet50V2, InceptionResNetV2, MobileNetV2, and DenseNet121, were applied to both individual and hybrid datasets. The results showed superior performance on the hybrid dataset, with VGG16, Xception, ResNet50V2, and DenseNet121 each achieving an accuracy of 99%. This consistent performance across the hybrid dataset highlights the robustness of these models in handling diverse data while maintaining high accuracy. To understand the models implicit behavior, explainable AI techniques were employed to illuminate their black-box nature. Specifically, LIME was used to enhance the interpretability of model predictions, especially in cases of misclassification, contributing to the development of reliable and interpretable AI-driven solutions for medical imaging.

Figures

Figures reproduced from arXiv: 2505.16028 by the authors.

Figure 1
Figure 1. Comparison of Different Datasets by Categories [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. A Flow Diagram of Our Proposed Model 4 Result Analysis The methodologies are executed on a computer system, with the best results selected for analysis. The system is equipped with an Intel Core i5 13400F processor, an NVIDIA GeForce RTX 3060 12GB GPU, and 16GB of DDR4 RAM. The models are implemented using Keras 3 , a high-level neural networks API, and TensorFlow 4 as the backend. To ensure robustness, each model r… view at source ↗
Figure 3
Figure 3. Some Sample Images of Misclassification by the Model [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

19 extracted references · 17 canonical work pages

  1. [14]

    Balanced augmented covid cxr dataset

    Mrinal Tyagi. Balanced augmented covid cxr dataset. https://www.kaggle.com/datasets/tr1gg3rtrash /balanced-augmented-covid-cxr-dataset , September 2022. Accessed on July 05, 2024

    work page 2022

  2. [2]

    Chest x-ray image

    Md Alamin Talukder. Chest x-ray image. https://data.mendeley.com/datasets/m4s2jn3csb/1, 2023

    work page 2023

  3. [15]

    Lung x-ray image

    Md Alamin Talukder. Lung x-ray image. https://data.mendeley.com/datasets/9d55cttn5h/1, 2023

    work page 2023

  4. [6]

    Curated dataset for covid-19 posterior-anterior chest radiography images (x-rays)

    UNAIS SAIT. Curated dataset for covid-19 posterior-anterior chest radiography images (x-rays). https: //data.mendeley.com/datasets/9xkhgts2s6/4, 2022

    work page 2022

  5. [1]

    Empowering covid-19 detection: Optimizing performance through fine-tuned efficientnet deep learning architecture

    Md Alamin Talukder, Md Abu Layek, Mohsin Kazi, Md Ashraf Uddin, and Sunil Aryal. Empowering covid-19 detection: Optimizing performance through fine-tuned efficientnet deep learning architecture. Computers in Biology and Medicine, 168:107789, 2024

    work page 2024

  6. [3]

    Covid-19 and pneumonia diagnosis from chest x-ray images using convolutional neural networks

    Muhab Hariri and Ercan Av¸ sar. Covid-19 and pneumonia diagnosis from chest x-ray images using convolutional neural networks. Network Modeling Analysis in Health Informatics and Bioinformatics , 12(1):17, 2023

    work page 2023

  7. [4]

    Deep learning in radiology: A transfer- learning based approach for the identification and classification of covid-19 and pneumonia in chest x-ray images

    Md Rabiul Hasan, Shah Muhammad Azmat Ullah, and Mehedi Hasan. Deep learning in radiology: A transfer- learning based approach for the identification and classification of covid-19 and pneumonia in chest x-ray images. In 2023 F ourth International Conference on Smart Technologies in Computing, Electrical and Electronics (ICSTCEE), pages 1–6. IEEE, 2023

    work page 2023

  8. [5]

    Comparative analysis of artificial intelligence for predicting covid-19 using diverse chest x-ray images

    Radhwan AA Saleh, Farid Al-Areqi, Zaid Al-Huda, and Mugahed A Al-antari. Comparative analysis of artificial intelligence for predicting covid-19 using diverse chest x-ray images. In 2023 3rd International Conference on Emerging Smart Technologies and Applications (eSmarTA), pages 1–7. IEEE, 2023

    work page 2023

Show all 19 references
  1. [7]

    Svd-clahe boosting and balanced loss function for covid-19 detection from an imbalanced chest x-ray dataset

    Santanu Roy, Mrinal Tyagi, Vibhuti Bansal, and Vikas Jain. Svd-clahe boosting and balanced loss function for covid-19 detection from an imbalanced chest x-ray dataset. Computers in Biology and Medicine , 150:106092, 2022

    2022

  2. [8]

    The performance evaluation of transfer learning vgg16 algorithm on various chest x-ray imaging datasets for covid-19 classification

    Andi Sunyoto, Yoga Pristyanto, Arief Setyanto, Fawaz Alarfaj, Naif Almusallam, and Mohammed Alreshoodi. The performance evaluation of transfer learning vgg16 algorithm on various chest x-ray imaging datasets for covid-19 classification. International Journal of Advanced Comput...

    2022

  3. [9]

    Covixnet: A novel and efficient deep learning model for detection of covid-19 using chest x-ray images

    Gaurav Srivastava, Aninditaa Chauhan, Mahesh Jangid, and Sandeep Chaurasia. Covixnet: A novel and efficient deep learning model for detection of covid-19 using chest x-ray images. Biomedical Signal Processing and Control, 78:103848, 2022

    2022

  4. [10]

    Explainable lung disease classification from chest x-ray images utilizing deep learning and xai

    Tanzina Taher Ifty, Saleh Ahmed Shafin, Shoeb Mohammad Shahriar, and Tashfia Towhid. Explainable lung disease classification from chest x-ray images utilizing deep learning and xai. arXiv preprint arXiv:2404.11428, 2024

    2024 arXiv

  5. [11]

    An explainable artificial intelligence model for multiple lung diseases classification from chest x-ray images using fine-tuned transfer learning

    Eram Mahamud, Nafiz Fahad, Md Assaduzzaman, SM Zain, Kah Ong Michael Goh, and Md Kishor Morol. An explainable artificial intelligence model for multiple lung diseases classification from chest x-ray images using fine-tuned transfer learning. Decision Analytics Journal, 12:100499, 2024

    2024

  6. [12]

    Lime-enabled investigation of convolutional neural network performances in covid-19 chest x-ray detection

    Eduardo Gasca Cervantes and Wai-Yip Chan. Lime-enabled investigation of convolutional neural network performances in covid-19 chest x-ray detection. In 2021 IEEE Canadian Conference on Electrical and Computer Engineering (CCECE), pages 1–6. IEEE, 2021

    2021

  7. [13]

    Comparative analysis of explainable artificial intelligence for covid-19 diagnosis on cxr image

    Joe Huei Ong, Kam Meng Goh, and Li Li Lim. Comparative analysis of explainable artificial intelligence for covid-19 diagnosis on cxr image. In 2021 IEEE International Conference on Signal and Image Processing Applications (ICSIPA), pages 185–190. IEEE, 2021

    2021

  8. [16]

    A study of cnn and transfer learning in medical imaging: Advantages, challenges, future scope

    Ahmad Waleed Salehi, Shakir Khan, Gaurav Gupta, Bayan Ibrahimm Alabduallah, Abrar Almjally, Hadeel Alsolai, Tamanna Siddiqui, and Adel Mellit. A study of cnn and transfer learning in medical imaging: Advantages, challenges, future scope. Sustainability, 15(7):5930, 2023

    2023

  9. [17]

    why should i trust you?

    Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. " why should i trust you?" explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1135–1144, 2016

    2016

  10. [18]

    Interpreting artificial intelligence models: a systematic review on the application of lime and shap in alzheimer’s disease detection

    Viswan Vimbi, Noushath Shaffi, and Mufti Mahmud. Interpreting artificial intelligence models: a systematic review on the application of lime and shap in alzheimer’s disease detection. Brain Informatics, 11(1):10, 2024

    2024

  11. [19]

    Investigating the effectiveness of interpretable cost-sensitive neural network for pneumonia detection contemplating data imbalance

    Ramisa Maliyat, Fatema Sultana, Sadia Sultana, Sajib Kumar Saha Joy, and Md Tanvir Rouf Shawon. Investigating the effectiveness of interpretable cost-sensitive neural network for pneumonia detection contemplating data imbalance. 2023. 10

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.