Reliability concern: This paper was retracted on 2026-03-17. See notice: 10.1038/s41598-026-44623-0.
Lung cancer is a leading cause of cancer-related death globally, and accurate diagnostic methods for early detection are crucial to enhance survival. The current diagnostic procedures and classical machine learning methods are helpful; however, modelling the structural variability of a lung nodule remains a long-standing and complex challenge. Although current deep learning approaches have achieved promising results, they still face challenges, including data scarcity (annotated datasets), bias (benign vs. malignant), overfitting (small sample sizes), and a black-box nature (non-interpretability), further hampering their use in clinical practice. These gaps also underscore the need for a robust, data-efficient, and explainable lung cancer detection framework. Thus, in this research, a GAN-augmented, attention-guided deep learning framework called LungGANDetectAI is proposed to address the shortcomings outlined above. It includes a GAN that generates synthetic CT images, alleviating class imbalance by augmenting samples of underrepresented classes with realistic images. Given the data size and imbalanced courses, this work proposed a classification task based on an attention model, GenAttnNet, that uses ResNet50 as its backbone and adds spatial and channel attention modules (SCAM) to enable the network to learn to focus on the most discriminative regions of the lung. Furthermore, Grad-CAM is applied to gradually reveal the clinical interpretability of the network’s decision-making. On the IQ-OTH/NCCD dataset, LungGANDetectAI achieves 98.4%, 98.2%, 98.3%, and 98.2% accuracy, precision, recall, and F1-score, respectively, only in malignant cases, effectively improving malignant case detection over baseline CNNs and state-of-the-art methods across extensive experiments. LungGANDetectAI is an accurate and interpretable decision-support tool that may serve as an early diagnostic aid for radiologists in lung cancer, leading to better patient-oriented services that improve outcomes.
LungGANDetectAI achieves 98.4%, 98.2%, 98.3%, and 98.2% accuracy, precision, recall, and F1-score, respectively, only in malignant cases
single small public dataset, selective malignant-only framing, and no external validation
LungGANDetectAI integrates class-balanced generative adversarial network (GAN) synthesis, attention-guided classification, and explainable visualisation
architecture and methods describe all three modules, though novelty is incremental
The results indicated consistent improvements in classification accuracy and recall (across all classes, especially those with low representation) in the presence of synthetic data
ablation is described qualitatively without complete quantitative reporting in the supplied text
Grad-CAM is applied to gradually reveal the clinical interpretability of the network’s decision-making
visual explanations are provided, but no radiologist agreement or lesion-ground-truth validation is shown
LungGANDetectAI is an accurate and interpretable decision-support tool that may serve as an early diagnostic aid for radiologists
clinical-aid framing outruns retrospective slice-level benchmark evidence
Derived from the full evaluation — not a separate score.
Strengths
The Methods state that all slices from a given patient are assigned to the same training, validation, or test subset. This is a meaningful safeguard for slice-based CT classification, where random slice-level splitting can overstate performance.
↳ Materials and methods, Dataset description and experimental splits
The framework links GAN-based class balancing, a ResNet50 classifier with spatial and channel attention modules, and Grad-CAM visualizations. This gives the paper a coherent engineering objective even if the components are not individually novel.
↳ Introduction; Materials and methods, LungGANDetectAI framework
The article targets early lung cancer detection from CT scans and frames the system as radiology decision support. The problem is concrete, active, and connected to identifiable clinical stakeholders.
↳ Abstract; Introduction
Limitations
All reported performance is based on IQ-OTH/NCCD, approximately 1190 slices from 110 patients, with no external validation. This sharply limits confidence that the reported malignant-case metrics would hold across scanners, institutions, or clinical populations.
↳ Materials and methods, Dataset description and experimental splits
The Synthetic data contribution section states that FID greater than 40 would indicate generated images are usable and realistic. Because lower FID conventionally indicates closer real-synthetic distributional similarity, this undermines confidence in the GAN-quality assessment.
↳ Materials and methods, Synthetic data contribution and utility
The Abstract emphasizes metrics “only in malignant cases” while the model is a three-class classifier, and the Methods contain inconsistent training descriptions and malformed notation. Garbled prose in the explainability section further reduces confidence in reporting precision.
↳ Abstract; Table 2; Eq. 6; Explainability and clinical interpretability integration
Withdrawn — shown for reference only
The paper has a coherent applied aim: address class imbalance, feature focus, and explainability in lung CT classification through a GAN-augmented attention pipeline. The patient-level split is a real methodological strength, but the evidence base is narrow because the evaluation uses one small public dataset and does not test external generalization. Confidence in the reported performance is further reduced by incomplete quantitative ablation reporting and the FID-threshold error in the synthetic-data section. The related work is extensive, but it often lists prior results rather than showing why this model advances beyond similar high-accuracy systems. Overall, the scores reflect a plausible proof-of-concept whose clinical and methodological claims exceed the validation provided.
Nabu’s assessment, alongside the field’s view.
Are you an author of this paper?
Limited2.8
Confidence mediumThe contribution is an applied integration of WGAN-GP augmentation, ResNet50-based spatial/channel attention, and Grad-CAM for lung CT classification. Its novelty is limited because the related work already documents many high-performing GAN, attention, hybrid, and explainability systems, and the paper does not quantitatively demonstrate its ablation advantage in the supplied results.
“integrates class-balanced generative adversarial network (GAN) synthesis, attention-guided classification, and explainable visualisation”
The methodology includes a meaningful patient-level 70/20/10 split and gives implementation details for preprocessing, WGAN-GP, ResNet50 attention modules, and training. The score is limited by the very small held-out patient count, no external validation, no cross-site robustness testing, qualitative rather than tabulated GAN ablation reporting, and the inverted FID threshold statement.
“the split is at the patient level”
The overall pipeline can be followed from problem framing to architecture and evaluation, but important parts of the manuscript are imprecise or corrupted. Selective malignant-only abstract metrics, repeated related-work descriptions, inconsistent end-to-end/separate training language, malformed equations, and garbled explainability prose reduce interpretive precision.
“only in malignant cases”
The paper cites extensive prior work across deep learning, GANs, attention, explainability, hybrid systems, and deployment barriers. Contextualization remains mostly descriptive, with limited synthesis explaining why the reported 98.4% result advances beyond cited studies reporting comparable or higher apparent performance.
“Gopinath et al. ... achieves high metrics (99.92% accuracy”
Concerns4 of 4 checks
The manuscript is broadly followable but contains material reporting and coherence issues that reduce confidence in the precision of the methods and claims. These concerns warrant attention but do not by themselves establish that the core results are invalid.
The study appears to use a named publicly available CT dataset and no prospective intervention is reported. Missing code, missing COI text, and lack of preregistration are recorded as openness or reporting limitations rather than conduct concerns for this design.
Flags: 1 declared / 5 total
93 references verified via PubMed, Crossref, or OpenAlex.
References not verified automatically are resolved manually within 24 hours.
This paper was retracted on 2026-03-17. See notice: 10.1038/s41598-026-44623-0.
Sources: Crossref (incl. Retraction Watch) ✓
Where this paper’s evidence sits on the path from initial observation to real-world use.
Readiness readiness is limited to a controlled proof-of-concept experiment. The paper itself notes that clinical use would require scan-level or volumetric inference, and it reports no prospective testing, workflow integration, regulatory pathway, or deployment study.
“to be clinically utilised, a scan-level or volumetric inference would be needed”
AI-generated, human-governed. Something look off? Contact us to request a review.
Keep reading
Other papers Nabu has evaluated blind, starting with the same field.
Han, Clancy, Arteaga-Reyes, et al. · 2026 · Clinical medicine
These findings support a nonatheromatous, intrinsic microvascular pathology, particularly segmental arteriolar disorganization, as the principal mechanism of lacunar stroke and cSVD.
TentativeXu, Ren, Wang, et al. · 2025 · Basic medicine
We observed increased forced vital capacity at the highest dosage with a mean change of +98.4 ml (95% confidence interval 10.9 to 185.9) for patients in the 60 mg rentosertib QD group
TentativeMaria Mironova, Bachir A. Taouli, Richard L. Ehman, et al. · 2026 · Clinical medicine
MRI proton density fat fraction (MRI-PDFF) is the most accurate noninvasive measure of steatosis (thresholds of 6.4%, 17.4% and 22.1% for grades ≥1, ≥2 and 3), and a ≥30% relative reduction is associated with histologic response.
Supported