Two-Way MR-Forest Based Growing Path Classification for Malignancy Estimation of Pulmonary Nodules

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Pulmonary nodules require accurate malignancy assessment to reduce false positives in lung cancer screening.
• Existing methods like MR-Forest and 3DDCNNs have limitations in dynamic analysis and false positive reduction.

Purpose of the Study:

• To propose a novel TMR-Forest framework for enhanced pulmonary nodule malignancy estimation.
• To improve false positive reduction (FPR) in lung nodule classification using pseudo-spatiotemporal features.

Main Methods:

• Utilized Mask R-CNN for region of interest (ROI) detection and feature classification.
• Implemented hierarchical attribute matching for ROI selection and growing path generation.
• Employed a two-stage counterfactual path elimination within a cascade forest structure.

Main Results:

• The TMR-Forest achieved a CPM score of 0.912 on 1034 scans.
• Demonstrated a 2.8% and 4.7% improvement in CPM score over MR-Forest and 3DDCNNs, respectively.
• Successfully reported more accurate malignancy labels for pulmonary nodules.

Conclusions:

• The TMR-Forest framework offers a dynamic approach to pulmonary nodule malignancy estimation.
• This method significantly enhances diagnostic accuracy and reduces false positives.
• The proposed technique shows superior performance compared to previous deep learning models.