Dual-threshold sample selection with latent tendency difference for label-noise-robust pneumoconiosis staging

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonary Medicine

Background:

• Pneumoconiosis staging on chest X-rays is challenged by progressive pair label noise (PPLN).
• Adjacent pneumoconiosis stages are often confused due to subtle lung opacities, degrading deep neural network performance.
• Existing methods struggle to effectively mitigate the impact of PPLN in medical image analysis.

Purpose of the Study:

• To enhance pneumoconiosis staging effectiveness by mitigating PPLN.
• To refine deep learning network architectures and sample selection mechanisms for noisy datasets.
• To improve the accuracy and reliability of AI-assisted pneumoconiosis diagnosis.

Main Methods:

• A novel multi-branch deep learning architecture with dual-threshold sample selection was developed.
• Auxiliary branches were integrated to learn progressive feature tendencies.
• A difference-based metric and iterative instance-specific thresholds were used for dynamic sample selection, partitioning data into 'clean' and 'hard' sets for differential loss treatment.

Main Results:

• The proposed method achieved superior performance metrics (accuracy: 90.92%, AUC: 94.64%) compared to state-of-the-art approaches under real-world PPLN.
• The model demonstrated reduced sensitivity to increasing rates of synthetic PPLN.
• Ablation studies confirmed the effectiveness of individual modules and hyperparameter impacts.

Conclusions:

• The developed method offers significant effectiveness and robustness against PPLN in pneumoconiosis staging.
• This approach can enhance diagnostic accuracy and physician confidence in identifying pneumoconiosis.
• The findings contribute to more reliable AI-driven diagnostic tools for occupational lung diseases.