Encrypted login | home

Program Information

A Radiomics Approach for Hyper-Dimensional Lung Function Mapping

no image available
K Lafata

K Lafata*, J Cai , C Kelsey , F Yin , Duke University Medical Center, Durham, NC


TH-AB-201-5 (Thursday, August 3, 2017) 7:30 AM - 9:30 AM Room: 201

Purpose: To investigate the potential correlation between pulmonary function and CT radiomic features extracted – both globally and locally – from the lungs.

Methods: First, to probe the relationship between radiomics data and global pulmonary function, 65 NSCLC patients were retrospectively identified with known DLCO(diffusion capacity of the lungs for carbon monoxide), and FEV1(forced expiratory volume in 1 second). The total lung volume was segmented from each patient’s planning-CT, from which 39 radiomic features were extracted as potential computational biomarkers. Each 39-dimensional biomarker was engineered to collectively capture intensity variations, fine-texture, and course-texture within the lungs. Unsupervised clustering was used to search for subsets of patients with similar radiomic profiles, where average FEV1 and DLCO values were calculated per-cluster. Next, each feature was treated as a finite impulse-response function and spatially-mapped throughout CT-segmented lungs, such that each biomarker dimension could be interpreted as its own intensity image. The relationship between this impulse-response and local pulmonary function was investigated on 3 patients by comparing the filtered images with available ⁶⁸Ga-PET perfusion-maps. Analysis was based on Mutual Information and reader-based structural CT-findings.

Results: Unsupervised clustering revealed subsets of patients with similar lung radiomic signatures. Statistically different mean FEV1 values were observed across three major clusters (p=0.001,0.002), indicating that patients with similar radiomic features also presented with comparable spirometry values. No such phenomenon was observed for DLCO(p>0.05). In particular, two texture features independently demonstrated high FEV1 signal: Long-Run-High-Gray-Level-Emphasis and Sum-Average(correlation coefficients of 0.73 and 0.64, respectively). Several texture-mapped images matched structural CT-findings, and Mutual Information was maximized between the Long-Run-High-Gray-Level-Emphasis and ⁶⁸Ga-PET images.

Conclusion: This study demonstrated a potential relationship between quantitative radiomic features and pulmonary function based on global and local lung measurements. Significant correlation was found between clustered radiomics data and FEV1, and hyper-textural filtering demonstrated potential for voxel-based pulmonary analysis.

Funding Support, Disclosures, and Conflict of Interest: Varian Medical Systems

Contact Email: