Semi-supervised adversarial model for benign-malignant lung nodule classification on chest CT.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate classification of lung nodules on chest CT is crucial for early lung cancer detection and patient survival.
• Deep convolutional neural networks (DCNNs) typically require extensive labeled data, which is often scarce in medical imaging applications due to annotation challenges.

Purpose of the Study:

• To develop a semi-supervised adversarial classification (SSAC) model for benign-malignant lung nodule classification using both labeled and unlabeled data.
• To enhance the model through multi-view knowledge-based collaborative learning (MK-SSAC) for comprehensive nodule characterization.

Main Methods:

• Proposed a semi-supervised adversarial classification (SSAC) model comprising an unsupervised reconstruction network (R) and a supervised classification network (C).
• Integrated learnable transition layers for adapting image representations from R to C.
• Extended the SSAC model to a multi-view knowledge-based collaborative learning framework (MK-SSAC) utilizing three SSACs for appearance, heterogeneity, and shape/texture characterization across nine planar views.

Main Results:

• The MK-SSAC model achieved a classification accuracy of 92.53% and an Area Under the Curve (AUC) of 95.81% on the LIDC-IDRI dataset.
• These results surpass the performance of existing lung nodule classification and semi-supervised learning methods.

Conclusions:

• The proposed MK-SSAC model demonstrates superior performance in classifying benign versus malignant lung nodules.
• This semi-supervised approach effectively addresses the challenge of limited labeled data in medical image analysis, paving the way for improved lung cancer diagnosis.