Quantitative texture analysis using machine learning for predicting interpretable pulmonary perfusion from...

Area of Science:

• Radiology
• Medical Imaging
• Machine Learning

Background:

• Pulmonary embolism (PE) diagnosis is critical but current methods like CT pulmonary angiography have limitations, especially for patients with contrast agent contraindications.
• Accurate and timely PE diagnosis is essential due to its life-threatening nature.

Purpose of the Study:

• To develop a quantitative texture analysis pipeline using machine learning (ML) on non-contrast CT scans.
• To discover intensity and textural features correlated with lung perfusion (Q) physiology and pathology.
• To synthesize voxel-wise Q surrogates to aid in PE diagnosis.

Main Methods:

• Retrospective collection of SPECT/CT scans from PE-suspected patients (internal and external cohorts).
• Extraction and two-stage selection of quantitative CT features, including prior-knowledge-driven and data-driven approaches.
• Training ML models for classification and synthesizing voxel-wise Q surrogates using selected feature maps.

Main Results:

• The optimal ML model achieved an AUC of 0.863 (internal) and 0.828 (external testing).
• A combination of 14 intensity and textural features effectively distinguished high- and low-functional lung regions.
• Moderate-to-high spatial agreement was observed between synthesized surrogates and reference Q-SPECT scans (median SCC 0.66, median DSCs 0.86/0.64).

Conclusions:

• Validated the feasibility of using quantitative texture analysis and ML for generating voxel-wise lung perfusion surrogates.
• This method offers a radiation-free, accessible alternative to functional lung imaging for pulmonary vascular diseases.
• The developed pipeline assists in PE diagnosis by providing non-contrast-based functional lung information.