Weakly-Supervised Segmentation-Based Quantitative Characterization of Pulmonary Cavity Lesions in CT Scans

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonary Medicine

Background:

• Pulmonary cavity lesions stem from diverse malignant and non-malignant lung diseases.
• Accurate diagnosis relies on recognizing characteristic morphological features on CT scans.
• Automated analysis of these lesions offers potential for improved clinical management.

Purpose of the Study:

• To develop a deep learning model for automatic detection, segmentation, and quantification of pulmonary cavity lesions.
• To quantitatively characterize cavity lesions using a weakly-supervised deep learning approach.
• To assess the model's performance in clinical diagnosis and treatment monitoring.

Main Methods:

• A weakly-supervised deep learning model (CSA2-ResNet) was developed.
• Lung parenchyma segmentation preceded feeding data into the deep neural network.
• Gradient-weighted class activation mapping and image processing were used for lesion visualization and segmentation.

Main Results:

• The model achieved high performance metrics: 98.48% accuracy, 96.80% precision, 97.20% specificity, 100% recall, and 98.36% F1-score.
• Demonstrated significant improvement over existing methods (P < 0.05).
• Quantitative morphological characterization yielded effective analysis results.

Conclusions:

• The proposed deep learning model offers a fast and effective method for diagnosing and monitoring pulmonary cavity lesions.
• AI-driven detection and quantitative analysis of these lesions can serve as potential indicators for diagnosis and dynamic monitoring.
• This approach has significant clinical and translational impact for patients with cavity lesions.