Saturday, November 21, 2026
Science
No Result
View All Result
  • Login
  • HOME
  • SCIENCE NEWS
  • CONTACT US
  • HOME
  • SCIENCE NEWS
  • CONTACT US
No Result
View All Result
Scienmag
No Result
View All Result
Home Science News

Deep Learning Covid-19 CT Models Show High Internal Accuracy but Generalization Limits

November 21, 2026
in Science News
Blake Davidson
By Blake Davidson Scienmag Editorial Profile - Data Science
Reading Time: 6 mins read
0
Deep Learning Covid-19 CT Models Show High Internal Accuracy but Generalization Limits

Deep Learning Covid-19 CT Models Show High Internal Accuracy but Generalization Limits

65
SHARES
587
VIEWS
Share on FacebookShare on Twitter
ADVERTISEMENT

Deep learning models developed for automated Covid-19 detection from computed tomography scans achieve high internal classification accuracy but experience notable performance drops when tested against unseen clinical patient cohorts from external institutions. A comprehensive international evaluation conducted across multiple multinational chest computed tomography datasets established that convolutional neural networks fine-tuned on public data frequently misclassify healthy individuals as positive when evaluating entirely independent hospital records. While artificial intelligence architectures reached diagnostic accuracy surpassing ninety-nine percent during internal cross-validation across standardized cohorts, evaluating external patient cohorts from Brazil and Algeria revealed persistent false-positive diagnoses. These empirical outcomes demonstrate that nominal algorithmic accuracy within training domains does not automatically guarantee generalization across heterogeneous clinical environments where imaging hardware, acquisition parameters, and institutional protocols vary substantially.

The deployment of computed tomography imaging became an essential diagnostic pillar during the pandemic because it provides superior sensitivity and morphological characterization compared to conventional radiography and standard antibody assays. Thoracic computed tomography scans reliably reveal characteristic viral pulmonary manifestations, specifically ground-glass opacities and pulmonary consolidation, which remain consistent across diverse viral variants. Furthermore, computed tomography can identify these pathological signs during the initial asymptomatic period and during later incubation phases when real-time reverse transcription polymerase chain reaction testing frequently demonstrates reduced diagnostic sensitivity. However, the manual interpretation of voluminous scan sequences demands substantial time from qualified radiologists, creating acute clinical bottlenecks in medically underserved regions experiencing specialist shortages during pandemic surges. This diagnostic burden prompted the rapid engineering of automated computer-aided systems utilizing deep learning algorithms to interpret radiological data swiftly.

The operational viability of automated radiological screening depends directly upon how convolutional neural networks respond to the physical variabilities inherent in medical imaging hardware. Scanners employ an array of adjustable acquisition parameters that determine the diagnostic quality, spatial resolution, and pixel intensity of resulting images. Tube current, measured in milliamperes, dictates radiation exposure, with higher currents generating elevated pixel intensity alongside increased image noise, whereas lower currents decrease exposure. Tube voltage, calibrated in peak kilovoltage, regulates beam penetration, where elevated voltage improves the visualization of dense bone structures and lower settings differentiate soft tissues. Concurrently, slice thickness modifies spatial resolution at the cost of total slice volume, pitch values balance scanning speed against resolution, and iterative reconstruction algorithms manage artifact suppression. Variations across these parameters frequently produce non-reproducible radiomics features, compromising diagnostic reliability across disparate clinical facilities.

To systematically assess whether existing diagnostic algorithms can maintain robustness despite these physical variations, researchers compiled and refined an expansive diagnostic repository drawn from seven international datasets. This collaborative collection integrated medical scans from thirteen countries, incorporating repositories such as CNCB, iCTCF, Negin Center, LIDC and IDRI, ITAC, MosMedData, and Radiopedia. Initial inspection revealed that 11,024 images, representing approximately 7.27 percent of the uncurated collection, lacked visible lung parenchyma because the slices fell entirely outside the functional lung field or captured non-informative anatomical margins. Because these irrelevant cross-sections cannot contribute to pulmonary assessment, the investigative team established a three-class classification scheme comprising Covid-19, Healthy, and Non-relevant slices. This deliberate filtering prevented non-diagnostic anatomical structures and scanner bed artifacts from distorting model training, ensuring algorithms evaluated genuine lung regions.

Constructing a balanced dataset for experimental modeling required implementing systematic lung segmentation and geometric data augmentation across the assembled collections. The researchers isolated the left and right lung boundaries to eliminate non-discriminating background elements, forcing algorithmic attention directly onto regions where pneumonia lesions physically manifest. To resolve representation imbalances between categories, data augmentation techniques applied horizontal reflections and multi-angle rotations at five, ten, fifteen, and twenty degrees to healthy and non-relevant slices. Following these transformations, the primary experimental library comprised 95,733 Covid-19 images, 95,944 healthy images, and 91,024 non-relevant images. In parallel, researchers curated DZ-CovidScan, an independent dataset comprising fifty-five verified Algerian computed tomography scans examined by a specialized physician, alongside fifty patient records selected from the BR-SARS-CoV-2 repository, reserving both collections exclusively for blind external generalization assessments.

The investigative framework systematically evaluated three prominent convolutional neural network families recognized for competitive image classification benchmarks: ResNet, DenseNet, and EfficientNet. The ResNet family, represented by ResNet18 and ResNet50, incorporates residual skip connections between successive network layers to facilitate back-propagation during gradient descent. The DenseNet architecture, tested across the DenseNet121, DenseNet161, and DenseNet201 configurations, connects each layer directly to all subsequent layers within four discrete DenseBlocks, optimizing feature reuse while maintaining lower computational overhead. In parallel, the EfficientNet series, encompassing versions b1, b3, b5, and b7, applies compound scaling methods to systematically balance network depth, layer width, and image resolution across variable feature map counts. By comparing these architectures under standardized experimental parameters, the study evaluated whether architectural differences fundamentally influence diagnostic sensitivity, parameter efficiency, and cross-dataset robustness.

Training the selected neural architectures followed a structured two-stage pipeline based on transfer learning and targeted architectural customization. In the initial stage, networks were initialized with pre-trained ImageNet weights and modified by replacing the standard thousand-unit output layer with a three-unit classification structure corresponding to the target classes. These baseline configurations underwent fine-tuning over twenty training epochs utilizing stochastic gradient descent with an initial learning rate of 0.00001, batch normalization with sizes ranging between 8 and 64, and 224-by-224 input resolutions. In the second stage, the top-performing networks with optimal parameter efficiency were redesigned by freezing their feature extraction layers and replacing the classification head. The customized classifier incorporated scaled exponential linear units to accelerate convergence, dropout regularizers set at a rate of 0.5, batch normalization, and a final SoftMax layer producing categorical pseudo-probabilities over ten training epochs.

Internal performance evaluation utilized an exhaustive fivefold cross-validation protocol across the seven primary training datasets, partitioning training and testing data under an exact four-to-one ratio per fold. During the initial fine-tuning stage, EfficientNet-b3 demonstrated superior predictive performance on early partitions, attaining 96.14 percent precision and 96.04 percent recall on fold one, alongside 98.98 percent precision and recall on fold two. DenseNet201 yielded the highest predictive metrics on fold three with 98.99 percent precision and fold four with 99.07 percent precision. Following the secondary transfer learning refinement, customized networks demonstrated marked performance gains, with modified DenseNet121 achieving peak single-fold values of 99.18 percent precision, 99.17 percent recall, 99.59 percent specificity, and 99.45 percent accuracy. Across all five cross-validation folds, modified DenseNet201 recorded the highest aggregate averages, maintaining 97.75 percent precision, 97.73 percent recall, 98.86 percent specificity, and 98.48 percent accuracy.

To establish whether exceptional cross-validation performance translates to genuine clinical utility, researchers subjected the six customized architectures to rigorous external validation against the unseen BR-SARS-CoV-2 and DZ-CovidScan datasets. External performance was evaluated on a per-patient basis across both cohorts using majority voting to synthesize slice-level determinations into an overall diagnostic determination for each individual scan sequence. Additionally, the Brazilian dataset underwent comprehensive per-image evaluation to assess individual slice accuracy across differing scanner outputs. Because these external repositories were strictly excluded from model development, parameter optimization, and cross-validation, they presented a realistic test environment encompassing distinct clinical acquisition protocols, reconstruction settings, and institutional hardware. This evaluation protocol directly examined whether models trained on aggregated public datasets could reliably classify pathological findings when confronted with novel, independent imaging sources.

External evaluation revealed a substantial divergence between internal cross-validation accuracy and real-world diagnostic reliability, particularly concerning false-positive diagnostic errors among non-infected cohorts. The customized models correctly identified every Covid-19 patient across both independent test datasets, demonstrating flawless sensitivity in recognizing confirmed positive cases. However, specificity deteriorated severely when evaluating healthy patient scans. The algorithms erroneously classified 14 out of 25 healthy subjects in the Brazilian repository and 6 out of 28 healthy subjects in the Algerian dataset as positive for Covid-19 infection. Within these misdiagnosed records, models misclassified between 1 and 20 individual slices per scan sequence. This pronounced error rate exposed a substantial generalization deficit, demonstrating that models trained on composite public repositories remain vulnerable to generating false alarms when confronted with imaging variations from new medical centers.

A notable contrast emerged between patient-level evaluation outcomes across the two external cohorts, reflecting significant operational differences in slice volume and scanning methodology. The Brazilian computed tomography scans averaged approximately twenty images per patient examination, whereas the Algerian scans averaged approximately three hundred images per patient sequence. Patient-level diagnostic evaluations yielded more conclusive results on the lower-volume Brazilian scans, indicating that acquiring extensive quantities of closely spaced slices introduces diagnostic redundancy without yielding corresponding gains in classification precision. On an individual slice basis within the Brazilian cohort, customized EfficientNet-b3 achieved the strongest performance, recording an area under the curve of 84.6 percent, a recall of 72.4 percent, and a precision of 95.83 percent, outperforming historical benchmarks while operating with fewer parameters than comparative ResNet50 baseline architectures.

To interpret the internal reasoning underlying these classification outcomes, researchers implemented Gradient-weighted Class Activation Mapping across true-positive, false-positive, and false-negative scans using customized DenseNet201. For true-positive Covid-19 scans, heat maps consistently highlighted unilateral and bilateral pulmonary lesions across superior, middle, and inferior lobes, confirming that the network correctly recognized legitimate pathological hallmarks. Conversely, visualization of false-positive scans revealed that the network misidentified benign anatomical abnormalities and benign pulmonary nodules as viral pneumonia, adopting these irrelevant features as false discriminating criteria. In false-negative examinations, the attention maps diffused across the entire lung parenchyma, failing to resolve sparse, subtle opacities. These interpretability findings confirm that deep learning diagnostics remain sensitive to irrelevant scan artifacts, highlighting the vital necessity of constructing comprehensive, multi-center international datasets to train genuinely robust clinical models.

Subject of Research: Generalization capabilities and diagnostic accuracy of deep learning models for Covid-19 detection using chest computed tomography scans

Article Title: Deep learning-based Covid-19 diagnosis: a thorough assessment with a focus on generalization capabilities

Article References: Hadj Bouzid, A. I., Berrani, S.-A., Yahiaoui, S., Belaid, A., Belazzougui, D., Djouad, M., Bensalah, K., Belbachir, H., Naïli, Q., Abdi, M. E.-H., & Tliba, S. (2024). Deep learning-based Covid-19 diagnosis: a thorough assessment with a focus on generalization capabilities. EURASIP Journal on Image and Video Processing, 2024(1), Article 40. https://doi.org/10.1186/s13640-024-00656-x

Image Credits: AI Generated

DOI: 10.1186/s13640-024-00656-x

Keywords: Covid-19 diagnosis, computed tomography, deep learning, convolutional neural networks, generalization capabilities, transfer learning, DenseNet, EfficientNet, ResNet, Grad-CAM, cross-validation, lung segmentation, radiomics, chest imaging, DZ-CovidScan, BR-SARS-CoV-2, false positives, medical image classification

Cite Scienmag News

Blake Davidson. (November 21, 2026). Deep Learning Covid-19 CT Models Show High Internal Accuracy but Generalization Limits. Scienmag. https://scienmag.com/deep-learning-covid-19-ct-models-show-high-internal-accuracy-but-generalization-limits/

Blake Davidson. "Deep Learning Covid-19 CT Models Show High Internal Accuracy but Generalization Limits." Scienmag, 21 November 2026, https://scienmag.com/deep-learning-covid-19-ct-models-show-high-internal-accuracy-but-generalization-limits/. Accessed 21 November 2026.

Blake Davidson. "Deep Learning Covid-19 CT Models Show High Internal Accuracy but Generalization Limits." Scienmag. November 21, 2026. https://scienmag.com/deep-learning-covid-19-ct-models-show-high-internal-accuracy-but-generalization-limits/

Tags: BR-SARS-CoV-2chest imagingcomputed tomographyconvolutional neural networksCOVID-19 diagnosiscross-validationdeep learningDenseNetDZ-CovidScanEfficientNetfalse positivesgeneralization capabilitiesGrad-CAMlung segmentationmedical image classificationradiomicsResNettransfer learning
Share26Tweet16
Previous Post

Hospital Pharmacists in China Face Systemic Barriers Despite Strong Research Motivation

Next Post

Tabular Federated Learning Updates Expose Private Records Under Gradient Inversion Attacks

Related Posts

Male Students in China Lead in Top Universities and STEM Majors
Science Education

Male Students in China Lead in Top Universities and STEM Majors

November 21, 2026
Storm Legacy Outweighs Cyclone Intensity in Coastal Forest Damage
Earth Science

Storm Legacy Outweighs Cyclone Intensity in Coastal Forest Damage

November 21, 2026
Pure Ethylene Glycol Coolant Cuts Simulated Metal Hydride Reactor Cycle Time
Technology and Engineering

Pure Ethylene Glycol Coolant Cuts Simulated Metal Hydride Reactor Cycle Time

November 21, 2026
Workplace Electronic Monitoring Curbs Job Crafting by Reducing Psychological Ownership
Science News

Workplace Electronic Monitoring Curbs Job Crafting by Reducing Psychological Ownership

November 21, 2026
Post-Exam Reflection Boosts Foundational Chemistry Gains but Leaves Complex Synthesis Unchanged
Science Education

Post-Exam Reflection Boosts Foundational Chemistry Gains but Leaves Complex Synthesis Unchanged

November 21, 2026
Multifunctional Polymer Framework Coating Blocks Ultraviolet Radiation and Enables Self-Healing Surfaces
Technology and Engineering

Multifunctional Polymer Framework Coating Blocks Ultraviolet Radiation and Enables Self-Healing Surfaces

November 21, 2026
Next Post
Tabular Federated Learning Updates Expose Private Records Under Gradient Inversion Attacks

Tabular Federated Learning Updates Expose Private Records Under Gradient Inversion Attacks

  • Mothers who receive childcare support from maternal grandparents show more optimized

    Mothers who receive childcare support from maternal grandparents show more parental warmth, finds NTU Singapore study

    27656 shares
    Share 11059 Tweet 6912
  • University of Seville Breaks 120-Year-Old Mystery, Revises a Key Einstein Concept

    1061 shares
    Share 424 Tweet 265
  • Bee body mass, pathogens and local climate influence heat tolerance

    682 shares
    Share 273 Tweet 171
  • Researchers record first-ever images and data of a shark experiencing a boat strike

    546 shares
    Share 218 Tweet 137
  • Groundbreaking Clinical Trial Reveals Lubiprostone Enhances Kidney Function

    531 shares
    Share 212 Tweet 133
Science

Embark on a thrilling journey of discovery with Scienmag.com—your ultimate source for cutting-edge breakthroughs. Immerse yourself in a world where curiosity knows no limits and tomorrow’s possibilities become today’s reality!

RECENT NEWS

  • Male Students in China Lead in Top Universities and STEM Majors
  • Storm Legacy Outweighs Cyclone Intensity in Coastal Forest Damage
  • Pure Ethylene Glycol Coolant Cuts Simulated Metal Hydride Reactor Cycle Time
  • Workplace Electronic Monitoring Curbs Job Crafting by Reducing Psychological Ownership

Categories

  • Agriculture
  • Anthropology
  • Archaeology
  • Athmospheric
  • Biology
  • Biotechnology
  • Blog
  • Bussines
  • Cancer
  • Chemistry
  • Climate
  • Earth Science
  • Editorial Policy
  • Marine
  • Mathematics
  • Medicine
  • Pediatry
  • Policy
  • Psychology & Psychiatry
  • Science Education
  • Science News
  • Social Science
  • Space
  • Technology and Engineering

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 5,151 other subscribers

© 2026 Scienmag - Science Magazine

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • HOME
  • SCIENCE NEWS
  • CONTACT US

© 2026 Scienmag - Science Magazine

Discover more from Science

Subscribe now to keep reading and get access to the full archive.

Continue reading