View Abstract

A2524 - Identifying Imaging Markers for Disease Progression in Longitudinal CTs of Patients with Idiopathic Pulmonary Fibrosis
Author Block: H. Prosch1, J. Pan1, J. Hofmanninger1, N. Sverzellati2, V. Poletti3, F. Paryer1, M. Holzer1, G. Langs1; 1Biomedical Imaging and Image-guided Therapy, Medical University of Vienna, Vienna, Austria, 2Radiology, University of Parma, Parma, Italy, 3Department of Thoracic Diseases, Morgagni-Pierantoni Hospital, Forli, Italy.
Normal 0 21 false false false DE X-NONE X-NONE /* Style Definitions */ table.MsoNormalTable {mso-style-name:""Normale Tabelle""; mso-tstyle-rowband-size:0; mso-tstyle-colband-size:0; mso-style-noshow:yes; mso-style-priority:99; mso-style-parent:""""; mso-padding-alt:0cm 5.4pt 0cm 5.4pt; mso-para-margin:0cm; mso-para-margin-bottom:.0001pt; text-align:justify; mso-pagination:widow-orphan; font-size:11.0pt; font-family:""Calibri"",sans-serif; color:black; mso-ansi-language:#0007;} Purpose To identify prognostic CT patterns in patients with idiopathic pulmonary fibrosis (IPF) using an unsupervised machine-learning approach. Specifically, we tested whether patterns could be identified that could be used as markers of disease progression in IPF. Materials and Methods A population of 106 IPF patients with at least one baseline and one follow-up CT scan were studied. CT scans per patient ranged from one to four scans, resulting in 695 scans included in the study. We identified patterns in the CT images by sub-dividing the lung into super voxels and extracting gray-level co-occurrence features at the centroid positions. Clustering in this feature space revealed 15 lung patterns that could be extracted with high stability across the population. Each voxel was then assigned in the entire data set to one of these clusters. The volume of each cluster relative to the entire lung was used as the signature of a lung. To identify prognostic markers in these signatures, we trained a random-forest classifier to predict for any pair of scans for one patient (scans A and B) if A was acquired prior to B or vice versa (overall 230 pairs). The classifier determined which features were most informative regarding the classification. Results The random forest marked four distinct clusters as predictive for the temporal sequence, enabling the identification of progression in follow-up scans. To determine whether the signature was predictive and stable, a four-fold cross-validation was performed on the data set. The classifier was trained on 3/4 of the patients, and predicted an A/B sequence on the remaining 1/4 of these patients. In the four-fold cross-validation experiment, the classifier correctly determined the sequence of scans for 80.35% of the cases. We identified four informative clusters. The ranking of the clusters was stable across the four folds, i.e., three clusters where among the top four most predictive clusters in all folds, and one cluster was in the top four for three of four folds. Discussion Using the described approach, we identified patterns that were markers of disease progression in lung CT data. The identification of these markers is data-driven, and enables the exploitation of complex patterns for the detection and quantification of progression.