MR-Forest: A Deep Decision Framework for False Positive Reduction in Pulmonary Nodule Detection.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Deep learning, particularly Convolutional Neural Networks (CNNs), has advanced automated pulmonary nodule detection accuracy.
• High computational and storage demands of large-scale neural networks (NNs) hinder widespread clinical adoption.

Purpose of the Study:

• To propose an alternative framework, Multi-ringed (MR)-Forest, for false positive reduction in pulmonary nodule detection.
• To address the resource-intensive nature of NN-based approaches for automated pulmonary nodule detection.

Main Methods:

• A novel multi-ringed scanning method extracts order ring facets (ORFs) from volumetric nodule models.
• Mesh-Local Binary Patterns (LBP) and mapping deformation estimate texture and shape features.
• Sliding and resampling ORFs generate multi-level feature volumes for cascaded prediction.

Main Results:

• The MR-Forest framework achieved a competitive Computerized Pulmonary Medicine (CPM) score of 0.865 in false positive reduction.
• Evaluated on 1034 scans from AH-LUTCM and LUNA16 datasets, demonstrating effectiveness.
• MR-Forest provides a balance between resource efficiency and high performance in pulmonary nodule detection.

Conclusions:

• MR-Forest presents a successful, resource-efficient solution for automated pulmonary nodule detection.
• The framework demonstrates strong competitiveness against state-of-the-art methods in false positive reduction.
• MR-Forest's general architecture is adaptable for other 3D medical imaging analysis tasks involving spheroidal expansion.