Feature-shared adaptive-boost deep learning for invasiveness classification of pulmonary subsolid nodules in CT...

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Pulmonary nodule classification

Background:

• Subsolid pulmonary nodules require accurate invasiveness assessment for clinical management.
• Classifying nodules into atypical adenomatous hyperplasia (AAH), adenocarcinoma in situ (AIS), minimally invasive adenocarcinoma (MIA), and invasive adenocarcinoma (IAC) is crucial.
• Automating invasiveness determination from CT scans is challenging due to data limitations and nodule variability.

Purpose of the Study:

• To develop a deep learning strategy for accurate invasiveness classification of subsolid pulmonary nodules.
• To address challenges of insufficient training data, interclass similarity, and intraclass variation in nodule classification.
• To guide treatment planning through automatic determination of nodule invasiveness.

Main Methods:

• A two-stage deep learning strategy: prior-feature learning followed by adaptive-boost deep learning.
• Utilizing multiple 3D convolutional neural network (CNN)-based weak classifiers within an adaptive-boost framework.
• Employing prior-feature learning to reduce computational load by sharing CNN layers among weak classifiers, integrating them into a single network.

Main Results:

• Achieved an AUC of 81.3% in binary classification of 1357 nodules (765 noninvasive, 592 invasive).
• Outperformed three experienced chest imaging specialists in accuracy.
• Demonstrated superior performance compared to non-ensemble deep learning methods on an additional 200 nodules across all categories (AAH, AIS, MIA, IAC).

Conclusions:

• Adaptive-boost deep learning significantly enhances invasiveness classification of pulmonary subsolid nodules in CT images.
• Prior-feature learning effectively reduces the computational size of deep models.
• The trained models show potential as effective lung cancer screening tools and the strategy is extensible to other 3D medical image classification tasks.