Robust Stochastic Neural Ensemble Learning With Noisy Labels for Thoracic Disease Classification.

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Chest radiography is crucial for diagnosing thoracic diseases like pneumonia.
• Deep learning models for computer-aided diagnosis (CADx) of thoracic diseases are widely used.
• Performance of deep learning models degrades with noisy labels common in real-world radiology.

Purpose of the Study:

• To develop a robust framework for thoracic disease diagnosis from chest X-rays.
• To address the challenge of performance degradation caused by noisy labels in deep learning models.
• To enhance the reliability of computer-aided diagnosis systems in clinical practice.

Main Methods:

• A novel stochastic neural ensemble learning (SNEL) framework was proposed.
• The method involves constructing model ensembles and designing noise-robust loss functions.
• A fast neural ensemble method collecting parameters simultaneously across instances and optimization trajectories was developed.
• A novel loss function optimizing robust and diversity measures for ensembles was designed.

Main Results:

• The SNEL method demonstrated superior performance compared to competing methods on three public chest X-ray datasets.
• Experimental results validated the effectiveness and robustness of SNEL in learning from noisy labels.
• The proposed framework successfully mitigates performance degradation associated with annotation biases.

Conclusions:

• The SNEL framework offers a robust solution for thoracic disease diagnosis using chest X-rays with noisy labels.
• This approach enhances the reliability of deep learning-based CADx systems in practical radiology.
• The study highlights the potential of ensemble learning and noise-robust loss functions for medical imaging AI.