COVID-19 Interpretable Diagnosis Algorithm Based on a Small Number of Chest X-Ray Samples

Area of Science:

• Artificial Intelligence in Medicine
• Medical Imaging Analysis
• Infectious Disease Diagnostics

Background:

• COVID-19 diagnosis using chest X-rays (CXRs) was initially challenging due to scarce data.
• The integration of artificial intelligence (AI) with medical diagnosis offers advanced solutions.
• Limited availability of COVID-19 CXR data hindered early research efforts.

Purpose of the Study:

• To develop an interpretable AI model for COVID-19 diagnosis from CXRs.
• To explore pathological characteristics of COVID-19 CXRs using AI interpretability.
• To address the challenge of limited sample sizes in medical AI model training.

Main Methods:

• Utilized AI interpretability analysis to understand CXR pathological features.
• Expanded the dataset using data augmentation to prevent overfitting.
• Employed transfer learning with custom output layers for training on limited samples.
• Compared four pre-trained models across three output layers and evaluated data augmentation effects.

Main Results:

• Achieved a high accuracy of 99.23% and a recall rate of 98%.
• Visual results from CXR interpretability analysis were successfully generated.
• The developed AI network demonstrated high generalization and a lightweight architecture.

Conclusions:

• The interpretable AI algorithm offers a robust and efficient method for COVID-19 diagnosis.
• The model's lightweight and generalizable nature allows rapid application to other data-scarce urgent tasks.
• AI interpretability analysis opens new avenues for enhancing medical diagnostic capabilities.