Automated pulmonary nodule classification from low-dose CT images using ERBNet: an ensemble learning approach.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate lung nodule detection is crucial for early lung cancer diagnosis.
• Low-dose CT (LDCT) imaging reduces radiation exposure but can compromise image quality, impacting nodule analysis.
• Developing robust deep learning models for diverse CT image qualities is essential.

Purpose of the Study:

• To develop and evaluate deep learning methods for classifying lung lesions as nodules or non-nodules across varying CT image doses and qualities.
• To compare the performance of dedicated models for LDCT versus a generalizable ensemble model.
• To optimize nodule classification accuracy using 3D convolutional neural networks (CNNs).

Main Methods:

• Utilized the Lung Nodule Analysis 2016 challenge dataset, generating LDCT images at 10%, 20%, 40%, and 60% dose levels from full-dose CT (FDCT) images.
• Developed five distinct 3D CNNs for nodule classification on both LDCT and FDCT images.
• Created an ensemble model to generalize across different dose levels and evaluated models on 800 samples (400 nodules, 400 non-nodules).

Main Results:

• A general model achieved 97.0% accuracy on FDCT images but performed poorly (60% accuracy) on LDCT images.
• Dedicated LDCT models showed significant performance improvements: 90.0% (10% dose), 91.1% (20%), 92.7% (40%), and 93.8% (60%).
• The ensemble model achieved 95.0% accuracy when tested on a combination of various dose levels, demonstrating generalizability.

Conclusions:

• Dedicated deep learning models are necessary for accurate lung nodule classification in LDCT images.
• An ensemble 3D CNN classifier effectively analyzes CT images across different dose levels and qualities.
• The developed models offer a promising approach for improving lung nodule detection in routine clinical practice with reduced radiation exposure.