MSCS-DeepLN: Evaluating lung nodule malignancy using multi-scale cost-sensitive neural networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate identification of malignant lung nodules in computed tomography (CT) screening is crucial for early lung cancer detection and patient cure.
• Deep learning shows promise but faces challenges with small datasets leading to overfitting and category imbalance.

Purpose of the Study:

• To propose MSCS-DeepLN, a method that evaluates lung nodule malignancy while addressing small dataset and data imbalance issues.
• To improve the accuracy and reliability of deep learning models for lung nodule malignancy identification.

Main Methods:

• Employs three light models with 3D convolutional neural networks (CNNs) as backbones to extract features and preserve spatial heterogeneity.
• Utilizes multi-scale input from CT images for sub-networks to learn multi-level contextual features.
• Incorporates an AUC approximation as a penalty term to handle data imbalance, ensuring equal contribution from positive and negative classes.

Main Results:

• Achieved state-of-the-art results on the LIDC-IDRI dataset.
• Demonstrated effective performance in addressing challenges of small datasets and category imbalance.
• Validated performance on a new dataset from a tertiary hospital, annotated via biopsy-based cytological analysis.

Conclusions:

• MSCS-DeepLN effectively evaluates lung nodule malignancy and overcomes common deep learning challenges in medical imaging.
• The method shows significant potential for clinical practice in early lung cancer diagnosis.
• Future work may involve further validation and integration into clinical workflows.