3D deep learning based classification of pulmonary ground glass opacity nodules with automatic segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate classification of Ground-Glass Lung Nodules (GGNs) into subtypes like atypical adenomatous hyperplasia (AAH), adenocarcinoma in situ (AIS), minimally invasive adenocarcinoma (MIA), and invasive adenocarcinoma (IAC) is crucial for effective lung cancer therapy.
• Diagnostic CT images are key for GGN classification, but subtle differences between subtypes pose challenges for automated analysis.

Purpose of the Study:

• To develop and evaluate a joint deep learning model that integrates nodule segmentation with classification to enhance the diagnostic performance for pulmonary GGNs.
• To investigate the efficacy of a cascade architecture where a segmentation network preprocesses CT data, generating an attention map to guide the classification network.

Main Methods:

• A novel cascade deep learning architecture was proposed, comprising a segmentation network followed by a classification network.
• The segmentation network acts as a trainable preprocessing module, generating a classification-guided 'attention' weight map from raw CT data.
• The model was evaluated on four clinically significant nodule classification tasks using metrics including Accuracy, Average F1 Score, Matthews Correlation Coefficient (MCC), and Area Under the ROC Curve (AUC).

Main Results:

• The proposed joint deep learning model demonstrated superior performance compared to baseline models across all four diagnostic classification tasks.
• The integration of segmentation-guided attention maps significantly improved the classification accuracy and reliability of GGN diagnosis.
• Experimental results confirmed the effectiveness of the cascade architecture in enhancing lesion classification by focusing on relevant nodule features.

Conclusions:

• The proposed joint deep learning model, incorporating a segmentation-guided attention mechanism, offers a significant advancement in the automated classification of pulmonary GGNs.
• This approach provides a more accurate and reliable method for differentiating GGN subtypes, aiding in better therapeutic strategy selection for lung cancer patients.
• The study highlights the potential of trainable preprocessing modules within deep learning frameworks for improving diagnostic performance in medical imaging analysis.