View Abstract

A6392 - Unsupervised Discovery of Spatial Informed Lung Texture Pattern (sLTP) for Pulmonary Emphysema Subtypes: The Subpopulations and Intermediate Outcome Measures in COPD Study (SPIROMICS)
Author Block: J. Yang1, E. D. Angelini2, P. P. Balte3, E. A. Hoffman4, J. H. Austin3, B. M. Smith3, E. R. Bleecker5, R. P. Bowler6, C. B. Cooper7, D. Couper8, M. T. Dransfield9, M. K. Han10, N. N. Hansel11, F. J. Martinez12, R. Paine13, J. D. Schroeder13, P. Woodruff14, V. E. Ortega15, A. F. Laine2, R. Barr16; 1Department of Biomedical Engineering, Columbia University, New York, NY, United States, 2Department of Biomedical Engineering and Radiology, Columbia University, New York, NY, United States, 3Department of Medicine, Columbia University Medical Center, New York, NY, United States, 4Departments of Radiology, Medicine and Biomedical Engineering, University of Iowa, Iowa City, IA, United States, 5Department of Medicine, University of Arizona, Tucson, AZ, United States, 6Department of Medicine, National Jewish Health, Denver, CO, United States, 7Department of Medicine, University of California Los Angeles, Los Angeles, CA, United States, 8Department of Biostatistics, University of North Carolina, Chapel Hill, NC, United States, 9Lung Health Center, University of Alabama at Birmingham, Birmingham, AL, United States, 10Department of Medicine, University of Michigan, Ann Arbor, MI, United States, 11Department of Medicine, Johns Hopkins University, Baltimore, MD, United States, 12Department of Medicine, Cornell University, New York, NY, United States, 13Department of Medicine, University of Utah, Salt Lake City, UT, United States, 14Department of Medicine, University of California, San Francisco, San Francisco, CA, United States, 15Department of Medicine, Wake Forest University, Winston Salem, NC, United States, 16Department of Medicine and Epidemiology, Columbia University Medical Center, New York, NY, United States.
Rationale: Three standard emphysema subtypes (centrilobular, panlobular and paraseptal emphysema) were initially defined based upon small autopsy series and are currently used by radiologists for visual CT interpretation, albeit with modest inter-rater agreement. We tested unsupervised machine-learning approaches in a large study to discover spatially-informed lung texture patterns (sLTPs), which may represent novel emphysema subtypes, through integrating spatial and texture information and examined the clinical significance of these patterns.
Methods: The Subpopulations and Intermediate Outcome Measures in COPD Study (SPIROMICS) is a prospective study of 2,981 participants (never smokers and smokers with or without COPD) who underwent highly-standardized CT scanning. Scans were randomly divided into two sets to learn the sLTPs independently (1462/1460 scans per set, both with average age 63±9 years, 53% male, 76% Caucasian, BMI 28±5 kg/m2, and pack-years of 41±28 and 46±30). Geometrical mapping was used to encode voxel positions within lungs. Local regions of interest (ROIs) with percent emphysema (using -950 HU threshold with a probabilistic model) above 5% were sampled for training. Texture features based on textons and spatial features based on geometrical mapping were generated for individual ROIs and were used to cluster the ROIs into 100 clusters via fusing the feature distances. Graph partitioning was used to group similar clusters and to generate a final set of sLTPs. The discovered sLTPs were used to label the whole dataset and measure percentage of sLTPs (%sLTP) per scan. Linear regression was used to adjust for age, sex, race/ethnicity, smoking status, pack-years, COPD status, FEV1 and %emphysema.
Results: Ten distinct sLTPs were discovered in both training sets. ROIs belonging to individual sLTPs were visually homogeneous. For each training set, three sLTPs were predominantly distributed in apical regions, one sLTP was predominant in anterior regions and three sLTPs were predominant in posterior regions. Reproducibility of the discovered sLTPs was high, with a labeling overlap of 0.85 on test ROIs, and Spearman correlation of corresponding %sLTP above 0.95 for all pairs of sLTPs. Average %sLTP values over the population show distinct patterns for COPD compared with controls. Six out of ten sLTPs were independently associated with increased respiratory symptoms.
Conclusion: Unsupervised learning enabled discovery of new image-based emphysema sLTPs that are reproducible, have distinct texture and spatial patterns, are distinguishable from percent emphysema, and may represent novel emphysema subtypes.