Open research questions in COVID-19 diagnosis using AI

What the literature leaves open

• Based on the synthesized evidence, the following recommendations are proposed: For TB Program Implementation: (1) Prioritize CAD products (CAD4TB v7, qXR v3.x, Lunit INSIGHT CXR) that have met WHO TPP benchmarks in peer-reviewed, independently conducted clinical evaluations. (2) Mandate local threshold calibration exercises using datasets representative of the target screening population prior to programmatic deployment. (3) Develop tailored clinical pathways for high-risk subgroups (PLWH, persons with prior TB history) incorporating supplementary diagnostic tests. (4) Implement CAD as a component of integrated, multi-test screening algorithms rather than as a standalone diagnostic tool. For Research: (1) Conduct large-scale prospective studies evaluating CAD performance in paediatric populations with paediatric-specific reference standards. (2) Prioritize implementation science research evaluating real-world effectiveness, equity, and health system integration of CAD. (3) Investigate novel algorithmic approaches to improve performance in PLWH and persons with prior TB. (4) Conduct health economic analyses in diverse country contexts, including Indonesia, to inform national TB program investment decisions. For Policy and Governance: (1) Advocate for transparent, equitable CAD pricing models that do not penalize high-volume public health screening programs. (2) Establish regulatory frameworks for software version updates and performance re-evaluation. (3) Develop data governance standards protecting patient radiological data sovereignty in low- and middle-income countries. (4) Integrate CAD into Indonesia's national TB elimination strategy, given the country's status as the second highest TB-burden country globally. © The Indonesian Journal of General Medicine 271 Research Article Volume 43, Issue No.01 . 2026 ISSN : 3048-104X REFERENCES 1. World Health Organization. Global Tuberculosis Report 2023. Geneva: WHO; 2023. Available from: https://www.who.int/publications/i/item/9789240083851 2. Sung J, Kibirige J, Asege L, et al. Performance of universal and stratified computer-aided detection thresholds for chest x-ray-based tuberculosis screening: a cross-sectional, diagnostic accuracy study.

  Review of Computer-Aided Detection (CAD) Software for Tuberculosis on Chest X-Rays : A Systematic Review of Randomized Controlled Trial and Primary Studies · 2026 · DOI

• Explainable AI (XAI) methods such as Grad-CAM, Grad-CAM++, and Score-CAM address the interpretability barrier, yet systematic quantitative comparisons across multiple CNN architectures remain scarce.

  Parameter-efficient deep learning for pneumonia detection on chest X-rays: A comparative evaluation of explainable AI methods · 2026 · DOI

• C_LIO_LIMost prospective validation studies have been conducted in high-income countries using datasets demographically and technically similar to the training data; evidence from South Asian LMIC settings is scarce.

  Diagnostic accuracy of a DenseNet-121 deep learning algorithm for chest radiograph triage in health assessment applicants: a prospective shadow-mode validation study in Nepal · 2026 · DOI

• This research proposed a robust framework for H. pylori detection in endoscopic images using an optimized CNN model extended with metaheuristic optimization techniques. This study illustrated that implementing dynamic hyperparameter optimization can notably reduce time and computational costs, serving as a powerful approach to assist clinicians as a decision-support system. It also provides deployment metrics and explain- ability using Grad-CAM. The proposed TSOA and SCROA algorithms outperformed conventional metaheuris- tic approaches and other state-of-the-art methods. Statistical analysis confirmed the sig- nificance of these improvements across multiple performance metrics. Future studies will focus on evaluating the proposed algorithms on larger and more diverse datasets, extending the optimization and evaluating generalization performance. In addition, investigating privacy-preserving learning paradigms and federated learning Lewis et al. Discover Artificial Intelligence (2026) 6:600 Page 36 of 37 can enable secure model aggregation and facilitate collaborative training across multiple healthcare institutions without sharing sensitive patient data.

  A novel trigonometric subpopulation and sine cosine range optimization framework to classify Helicobacter pylori infection in a south Indian cross-sectional study · 2026 · DOI

• The dataset used in this study, although sufficient for initial experimentation, is relatively small and specific to pediatric patients from a single medical center. This limits the generalizability of the findings to broader populations and different age groups. Also, the dataset’s demographic homogeneity may not capture the variability in CXR images that could arise from different ethnicities, geographical locations, or adult populations. The inherent stochastic nature of deep learning training processes posed difficulties in consistently evaluating model performance. Variability in results necessitated the introduction of additional metrics to provide a more robust measure of performance. We ensured reproducibility and stability of the results across different runs by carefully managing random seeds and training conditions. We encountered significant computational demands when training deep learning models, especially with fine-tuning and multiple preprocessing techniques, and limited access to high-performance computing facilities slowed down the experimentation process. While several preprocessing techniques were applied, we did not explore all possible methods. There may be other techniques that could further enhance image quality and model performance. Additionally, some preprocessing methods might not be applicable to CXR images as they could alter the inherent nature of the data. For example, techniques could inadvertently obscure or modify critical diagnostic features, such as white shadows that are indicative of certain pathologies in CXRs. We evaluated the impact of preprocessing incorporating data augmentation, which is commonly used in practice to enhance model robustness. color distributions techniques without adjust that The study focused on EfficientNet B1 and did not extensively compare other state-of-the-art architectures. While EfficientNet is known for its performance, other models might offer different advantages. The decision to limit the comparison was influenced by processing load considerations and computational constraints, which restricted the scope of the study. The ROI selection process relied on an object detection model trained on a small subset of labeled data. This might introduce bias or errors if the object detection model is not generalizable to all types of CXR images. Also, images where the rib cage was not detected were included in the dataset without modifications, potentially affecting the consistency of the dataset.

  The Role of Weight Initialization and Preprocessing Techniques in Analyzing Chest X-ray Images with Deep Neural Networks: A Comparative Study · 2026 · DOI

• Consequently, these findings should be validated through comparative studies incorporating data from additional scanner manufacturers before broader clinical application can be considered. While this paper presents a novel statistical approach for identifying digital signatures of COVID-19 related lung consolidation patterns, an important limitation must be acknowledged.

  Covid-19 related Lung Consolidation’s Digital signature in CT scans · 2026 · DOI

• Although SARS-CoV-2 has been extensively studied from clinical, virological, and diagnostic perspectives, the problem of accurate automatic semantic segmentation of SARS-CoV-2 particles in electron microscopy images remains inadequately explored.

  Cascade Semantic Segmentation by a Convolutional Neural Network in Combination with Image Super-Euclidean Pixels Processing for SARS-CoV-2 Microscopy Images · 2026 · DOI

• Accurate diagnosis of neurological disorders is contingent upon advanced imaging modalities such as Magnetic Resonance Imaging (MRI), which commonly utilize sparse imaging techniques to reconstruct images from limited data, thus reducing storage and acquisition time.

  Multi-Class Neurological Disorder Prediction with Tensor Network Feature Engineering · 2026

• While recent studies in PAI have started to investigate subject-related confounders, the impact of hardware-related confounders remains unexplored, posing a critical risk for failure in multicentric deployment scenarios.

  Photoacoustic device fingerprints induce bias in deep learning models · 2026 · DOI

• Future work may investigate model compression techniques, efficient transformer variants, and lightweight architectures to further reduce computational cost and enable real-time deployment in clinical environments. Future research may explore multi-scale feature fusion, graph construction strategies beyond epsilon-radius connectivity, and clinician-in-the-loop validation to further assess interpretability and generalization across diverse endoscopic datasets.

  Grad-CAM based deep learning analytics for image-level colon disease classification based on graph neural networks and vision transformers · 2026 · DOI

• One of the main weaknesses of this study is that it uses only one dataset, which prevents the comparison of the model with different patient demographics. The existing data distribution is extremely biased towards adult groups, 10|Nwohiri et al. Period. Polytech. Elec. Eng. Comp. Sci. i.e., the fitted hyperparameters will not perfectly apply to pediatric patients. Moreover, given that the process of data division was done on an image-by-image basis but not on a patient-by- patient basis, two or more radiographs of the same patient can be found on the training and testing sets. External validation should take precedence in future work to tackle the limitations inherent in the use of single datasets. The optimized hyperparameter settings should be tested using independent X-ray data that are acquired from other hospitals to determine whether they are diag- nostic in real clinical setting. Scalability is another consideration because training deep learning models like InceptionV3 is resource- and time-intensive. Future research should focus on the use of hyperparameter tuning and less resource-intensive CNN architectures that train faster.

  Systematic Hyperparameter Optimization of Convolutional Neural Networks for Pneumonia Detection from Chest X-rays · 2026 · DOI

• The ResNet50 pre-trained architecture achieves identical 97.78% accuracy as CNN with F1-score 0.978, suggesting potential redundancy in the hybrid framework; ablation studies isolating the contribution of each component (CNN vs. ResNet50 vs. RNN vs. LSTM) and their interaction effects are needed to justify the four-model ensemble for lung disease recurrence prediction.

  Performance Analysis of a Hybrid Deep Learning Framework Integrating CNN, RNN, LSTM, and ResNet50 for Lung Disease Recurrence Prediction Using Chest X-Ray Images and Post-Recovery Clinical Data · 2026 · DOI

• The comparison with existing studies (Table IX) shows the framework achieves 97.78% on image classification versus 90-95% in prior work, but inconsistent evaluation metrics (accuracy vs. sensitivity/specificity/AUC) and different dataset compositions prevent direct performance benchmarking; standardized evaluation protocols for hybrid deep learning frameworks on chest X-ray and clinical data must be established.

  Performance Analysis of a Hybrid Deep Learning Framework Integrating CNN, RNN, LSTM, and ResNet50 for Lung Disease Recurrence Prediction Using Chest X-Ray Images and Post-Recovery Clinical Data · 2026 · DOI

• The LSTM model for temporal pattern modeling of lung disease recurrence operates on unspecified clinical features and visit sequences; the specific longitudinal variables (e.g., symptom progression markers, biomarker trends, treatment adherence), optimal sequence length, and temporal granularity (daily, weekly, monthly) that maximize LSTM predictive performance require systematic investigation.

  Performance Analysis of a Hybrid Deep Learning Framework Integrating CNN, RNN, LSTM, and ResNet50 for Lung Disease Recurrence Prediction Using Chest X-Ray Images and Post-Recovery Clinical Data · 2026 · DOI

• The proposed medical Decision Support System (DSS) for continuous post-recovery patient monitoring is conceptual; implementation requirements including real-time inference latency constraints, integration with Electronic Health Record systems, clinical workflow adaptation, and user interface design for the hybrid deep learning framework must be experimentally validated in actual clinical environments.

  Performance Analysis of a Hybrid Deep Learning Framework Integrating CNN, RNN, LSTM, and ResNet50 for Lung Disease Recurrence Prediction Using Chest X-Ray Images and Post-Recovery Clinical Data · 2026 · DOI

• The paper reports a 70-15-15 train-validation-test split but does not specify cross-validation strategies, class imbalance handling (particularly for the tuberculosis class noted as highest proportion), or robustness testing across different data distributions; these validation approaches must be implemented to confirm the 97.78% CNN/ResNet50 accuracy and 90% LSTM accuracy are not artifacts of the specific data split.

  Performance Analysis of a Hybrid Deep Learning Framework Integrating CNN, RNN, LSTM, and ResNet50 for Lung Disease Recurrence Prediction Using Chest X-Ray Images and Post-Recovery Clinical Data · 2026 · DOI

• The study used a relatively limited dataset from three clinical facilities in Gorontalo with unspecified total sample size; validation of the hybrid framework on larger, geographically diverse cohorts with multi-institutional chest X-ray images and post-recovery clinical data is necessary to establish generalizability across different lung disease populations and imaging protocols.

  Performance Analysis of a Hybrid Deep Learning Framework Integrating CNN, RNN, LSTM, and ResNet50 for Lung Disease Recurrence Prediction Using Chest X-Ray Images and Post-Recovery Clinical Data · 2026 · DOI

• The hybrid deep learning framework achieved 97.78% accuracy on chest X-ray images but only 90% accuracy on longitudinal clinical data using LSTM; the integration mechanism between CNN/ResNet50 spatial features and LSTM temporal features for multimodal recurrence prediction requires explicit fusion methodology and optimization to improve the clinical data component performance.

  Performance Analysis of a Hybrid Deep Learning Framework Integrating CNN, RNN, LSTM, and ResNet50 for Lung Disease Recurrence Prediction Using Chest X-Ray Images and Post-Recovery Clinical Data · 2026 · DOI

• The comparative performance table (Table IV) contains incomplete metrics across models (missing recall values for InceptionV3, VGG19, DenseNet201, MobileNet; missing precision for CNN [37]); standardized evaluation protocols and consistent metric reporting across all TB detection architectures are required for fair comparison.

  Enhancing Tuberculosis Detection from Chest X-Ray Images Using Deep Learning: Evaluating Multi-Architecture Performance and Efficiency · 2026 · DOI

• While the paper demonstrates VGG16 and MobileNetV2 performance on multi-source chest X-ray datasets, it does not evaluate model robustness across different imaging equipment manufacturers, radiograph quality variations, or geographic populations; large-scale multi-center validation with documented imaging device specifications and demographic stratification is needed.

  Enhancing Tuberculosis Detection from Chest X-Ray Images Using Deep Learning: Evaluating Multi-Architecture Performance and Efficiency · 2026 · DOI

• The custom CNN model exhibited overfitting during extended training despite achieving fast inference (~9-9.5 ms/image); the paper lacks investigation into regularization techniques, early stopping strategies, or architectural modifications specifically designed to prevent overfitting in lightweight TB detection CNNs without sacrificing speed.

  Enhancing Tuberculosis Detection from Chest X-Ray Images Using Deep Learning: Evaluating Multi-Architecture Performance and Efficiency · 2026 · DOI

• The development and implementation of the proposed multimodal data fusion and classification framework open several avenues for future research. These suggestions aim to extend the methodology, address the identified challenges and limitations, and enhance the overall efficacy of the system. ADVANCED DATA FUSION TECHNIQUES Extension: Investigating more sophisticated data fusion techniques, such as deep learning-based methods or advanced ensemble models, can further improve the integration of multimodal data. Exploring hybrid models that combine traditional machine learning algorithms with deep learning approaches may yield better fusion outcomes. Improvement Area: Enhanced fusion techniques could lead to more accurate and nuanced interpretations of the combined data, potentially improving diagnostic precision and reliability. CROSS-MODAL DATA AUGMENTATION Extension: Future research could explore cross-modal data augmentation strategies to address issues related to data scarcity or imbalance in one or more modalities. Generating synthetic data based on existing modalities could enhance the robustness and generalizability of the model. Improvement Area: This would allow the model to be trained on a more diverse and comprehensive dataset, improving its performance across a wider range of clinical scenarios. EXPLAINABLE AI IN MEDICAL DIAGNOSTICS Extension: Incorporating techniques from the field of explainable AI (XAI) to improve the interpretability of the model's decision-making process. This could involve developing methods to visualize and explain how the model arrives at a particular diagnostic conclusion. Improvement Area: Enhanced interpretability will not only increase trust among healthcare professionals but also provide valuable insights into the diagnostic process, potentially informing clinical decision-making. REAL-TIME PROCESSING CAPABILITIES Extension: Adapting the framework for real-time processing and diagnostics could be a significant advancement. This involves optimizing the models for faster computation without sacrificing accuracy. Improvement Area: Real-time diagnostic capabilities would make the framework suitable for emergency medical situations where rapid decision-making is crucial. DIVERSE POPULATION GENERALIZABILITY Extension: Conducting studies on more diverse datasets that encompass a wider range of demographics, geographic locations, and medical conditions. This would test the model's generalizability and performance across different patient populations. Improvement Area: Enhancing the model's applicability and accuracy across diverse patient groups, thereby reducing potential diagnostic disparities.

  Harnessing the Power of Multimodal Data: Medical Fusion and Classification · 2024 · DOI

• Furthermore, I argue that the reliability of AI methodologies such as deep neural networks-which are at the center of this argument-is something that has not yet been established, and doing so faces fundamental challenges.

  Should we replace radiologists with deep learning? Pigeons, error and trust in medical AI · 2021 · DOI

• Future work will focus on replacing projected ablation values with full measured logs, improving rare-class stability, validating the model on external chest X-ray datasets, and adding clinical reader studies or calibrated uncertainty analysis before practical deployment.

  CDO-VIB DenseNet: An Improved DenseNet-121 Network for Multi-Label Chest X-ray Disease Classification · 2026 · DOI

• Limited data, unlimited potential: A study on vits augmented by masked autoencoders. Nevertheless, further expansion of the dataset and validation across different imaging protocols are warranted to confirm the generalizability of the proposed approach.

  A multi-task masked autoencoder with GAN-based augmentation for PD-L1 prediction from chest CT images · 2026 · DOI

Most-cited papers in COVID-19 diagnosis using AI

• Evaluation metrics and statistical tests for machine learning · Scientific Reports · 2024 · 1,035 citations
• Iterative enhancement fusion-based cascaded model for detection and localization of multiple disease from CXR-Images · Expert Systems with Applications · 2024 · 387 citations
• A Comprehensive Systematic Review of YOLO for Medical Object Detection (2018 to 2023) · IEEE Access · 2024 · 259 citations
• Deep learning based multimodal biomedical data fusion: An overview and comparative review · Information Fusion · 2024 · 232 citations
• YOLOv1 to YOLOv10: The Fastest and Most Accurate Real-time Object Detection Systems · APSIPA Transactions on Signal and Information Processing · 2024 · 147 citations
• Advancements and Prospects of Machine Learning in Medical Diagnostics: Unveiling the Future of Diagnostic Precision · Archives of Computational Methods in Engineering · 2024 · 124 citations
• Efficient pneumonia detection using Vision Transformers on chest X-rays · Scientific Reports · 2024 · 107 citations
• Raising the Bar of AI-generated Image Detection with CLIP · 2024 · 105 citations
• CTBViT: A novel ViT for tuberculosis classification with efficient block and randomized classifier · Biomedical Signal Processing and Control · 2024 · 105 citations
• Secure and Transparent Lung and Colon Cancer Classification Using Blockchain and Microsoft Azure · Advances in respiratory medicine · 2024 · 104 citations

Most recent work

• Performance Analysis of a Hybrid Deep Learning Framework Integrating CNN, RNN, LSTM, and ResNet50 for Lung Disease Recurrence Prediction Using Chest X-Ray Images and Post-Recovery Clinical Data · Engineering, Technology & Applied Science Research · 2026
• Research on lung nodule detection in X-ray plain films based on improved YOLOv12 model · Scientific Reports · 2026
• Hybrid deep learning metaheuristic ensemble framework for enhanced COVID-19 classification using ResNet and particle swarm optimization · Discover Artificial Intelligence · 2026
• Deep Learning–Based Automated Diagnostic Charting on Panoramic Radiography: Comparison of YOLOv11 and YOLOv12 · Odontology · 2026
• A machine learning-based prediction model for treatment efficacy in smear and/or chest X-ray positive tuberculosis patients · BMC Infectious Diseases · 2026
• Combining semantic CNN and HOG texture with LightGBM for explainable tuberculosis detection · Discover Artificial Intelligence · 2026
• A hybrid feature selection framework collaborating image processing, evolutionary intelligence, and machine learning for covid-19 disease classification · Network Modeling Analysis in Health Informatics and Bioinformatics · 2026
• Multi-label medical diagnosis using spatial-disease feature condensation and Kolmogorov–Arnold layers · Physics in Medicine & Biology · 2026
• Enhancing Tuberculosis Detection from Chest X-Ray Images Using Deep Learning: Evaluating Multi-Architecture Performance and Efficiency · Engineering, Technology & Applied Science Research · 2026
• Medical Image Classification for Pneumonia Detection Using Deep Learning · INTERNATIONAL JOURNAL OF SCIENTIFIC RESEARCH IN ENGINEERING AND MANAGEMENT · 2026

Related topics in Medicine