[home] [Personal Program] [Help]

---

Session: Poster session II
Session starts: Thursday 24 January, 16:00

---

Yeshaswini Nagaraj (University of Groningen, University Medical Center Groningen)
Jiapan Guo (University of Groningen, University Medical Center Groningen)
Monique D. Dorrius (University of Groningen, University Medical Center Groningen)
Mieneke Rook (University of Groningen, University Medical Center Groningen)
Qiong Li (University of Groningen, University Medical Center Groningen)
Rozemarijn Vliegenthart (University of Groningen, University Medical Center Groningen)
Matthijs Oudkerk (University of Groningen, University Medical Center Groningen)
Peter . M.A van Ooijen (University of Groningen, University Medical Center Groningen)

Abstract:
Emphysema is part of the chronic obstructive pulmonary disease (COPD) spectrum and known for its high prevalence and mortality rate worldwide. The appearing lung tissue destruction can be detected on Low-Dose Computed Tomography (LDCT) as decreased lung density and architectural changes. However, the latter is difficult to quantify so far. In this study, generalised matrix learning vector quantisation (GMLVQ) classifier was employed to discriminate between emphysematous and non-emphysematous lung tissue in LDCT by using radiomics features. 64 LDCT scans were randomly selected from a large lung cancer screening trial. Independently, three radiologists selected a total of 419 free form Regions of Interest (ROIs) from these scans. Every ROI consisted of three consecutive non-overlapping 2D slices. 300 ROIs were classified as emphysematous regions by the radiologists. During the multi-randomised training, each time 100 emphysematous regions were compared to 100 non-emphysematous regions. ROIs were used to extract radiomics features, which were ranked based on relevance factors from GMLVQ. We used feature selection techniques to decrease redundancy between features before performing the classification. Cross-validation was done using leave-one out procedure for each training set and number of feature selected. We extracted 1008 radiomics features including Gray-Level Co-Occurrence Matrix (GLCM), Gray-Level Run-length Matrix (GLRL), first order histogram features, shape descriptors and image filters. The GMLVQ classifier with the relevance feature ranking, which can be related to pairwise correlations between features, resulted in a maximal area under the curve of 0.95 for discrimination between emphysematous and non-emphysematous lung tissue. GMLVQ using radiomic features can discriminate between lung tissue with and without emphysema. Potentially, it can be a valuable tool for characterization of emphysema.