Development of Debiasing Technique for Lung Nodule Chest X-ray Datasets to Generalize Deep Learning Models.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Lung cancer screening is limited by accessibility issues for rural populations.
• Current screening methods face challenges in cost, speed, and privacy.

Purpose of the Study:

• To develop a pre-processing pipeline for chest X-ray images to improve deep learning-based lung nodule detection.
• To enhance the accuracy and generalization of lung cancer screening models for remote areas.

Main Methods:

• A pre-processing pipeline was developed, including histogram equalization, lung field segmentation, cropping, and rib/bone suppression.
• A deep learning model for nodule detection was trained using a pruning mechanism on a public lung nodule X-ray dataset.
• Federated deep learning principles were considered to ensure data privacy and prevent model bias.

Main Results:

• The pre-processing pipeline successfully debiased chest X-ray images, enhancing model classification and generalization.
• The deep learning models demonstrated successful generalization on an independent dataset.
• An external generalization accuracy of 89% was achieved for lung nodule detection.

Conclusions:

• The proposed deep learning algorithm effectively detects lung nodules by mitigating image noise and confounding variables.
• This approach enables the development of low-cost, accessible deep learning systems for widespread lung cancer screening.
• The findings support the potential for mobile and private screening solutions for early lung cancer diagnosis.