Related Experiment Video
Updated: Oct 16, 2025
Lung CT Segmentation to Identify Consolidations and Ground Glass Areas for Quantitative Assesment of SARS-CoV Pneumonia
Published on: December 19, 2020
Current limitations to identify covid-19 using artificial intelligence with chest x-ray imaging (part ii). The
José Daniel López-Cabrera1, Rubén Orozco-Morales2, Jorge Armando Portal-Díaz2
1Centro de Investigaciones de La Informática, Facultad de Matemática, Física y Computación, Universidad Central "Marta Abreu" de Las Villas, Villa Clara, Santa Clara, Cuba.
Artificial intelligence models for COVID-19 detection using chest X-rays show high sensitivity but often rely on "shortcut learning." These models frequently identify irrelevant image regions, limiting their clinical applicability and generalizability.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- COVID-19 pandemic spurred AI research for disease identification via radiological images.
- Automatic classification methods report high sensitivity, exceeding human specialists.
- Radiology societies caution against using imaging alone due to unspecific patterns overlapping with other viral pneumonias.
Purpose of the Study:
- Evaluate the robustness and generalizability of AI, deep learning, and computer vision for COVID-19 detection using chest X-rays.
- Alert researchers to the issue of "shortcut learning" in these models.
- Provide recommendations to identify shortcut learning in COVID-19 classification models.
Main Methods:
- Review of papers employing explainable artificial intelligence (XAI) methods.
- Evaluation of external validation sets to determine model generalizability.
- Consideration of traditional computer vision methods for COVID-19 classification.
Main Results:
- Key image regions contributing to classification often lie outside the lung area, suggesting reliance on non-pathological features.
- Model effectiveness significantly decreases when evaluated on external datasets, indicating poor generalizability.
- Existing models frequently exhibit shortcut learning, compromising their clinical utility.
Conclusions:
- Current AI models for COVID-19 detection via chest X-rays are often affected by shortcut learning.
- The identified shortcuts limit the reliability and generalizability of these models in clinical settings.
- Further research is needed to develop robust and clinically applicable AI diagnostic tools.
