Computer-aided diagnosis of pulmonary diseases using x-ray darkfield radiography. (17th November 2015)

• Record Type: Journal Article
• Authors: Einarsdóttir, Hildur; Yaroshenko, Andre; Velroyen, Astrid; Bech, Martin; Hellbach, Katharina; Auweter, Sigrid; Yildirim, Önder; Meinel, Felix G; Eickelberg, Oliver; Reiser, Maximilian; Larsen, Rasmus; Ersbøll, Bjarne Kjær; Pfeiffer, Franz
• Is Part Of: Physics in medicine & biology. Volume 60:Number 24(2015:Dec.)
• Page Start: 9253
• Page End: 9268
• Publication Date: 2015-11-17
• Subjects: X-ray radiography -- dark-field imaging -- lung segmentation -- active appearance model -- pulmonary disease -- Grating based interferometry; Biophysics -- Periodicals; Medical physics -- Periodicals; 610.153
• DOI: 10.1088/0031-9155/60/24/9253
• Languages: English
• ISSNs: 0031-9155

Abstract

Abstract: In this work we develop a computer-aided diagnosis (CAD) scheme for classification of pulmonary disease for grating-based x-ray radiography. In addition to conventional transmission radiography, the grating-based technique provides a dark-field imaging modality, which utilizes the scattering properties of the x-rays. This modality has shown great potential for diagnosing early stage emphysema and fibrosis in mouse lungs in vivo . The CAD scheme is developed to assist radiologists and other medical experts to develop new diagnostic methods when evaluating grating-based images. The scheme consists of three stages: (i) automatic lung segmentation; (ii) feature extraction from lung shape and dark-field image intensities; (iii) classification between healthy, emphysema and fibrosis lungs. A study of 102 mice was conducted with 34 healthy, 52 emphysema and 16 fibrosis subjects. Each image was manually annotated to build an experimental dataset. System performance was assessed by: (i) determining the quality of the segmentations; (ii) validating emphysema and fibrosis recognition by a linear support vector machine using leave-one-out cross-validation. In terms of segmentation quality, we obtained an overlap percentage (Ω) 92.63 ± 3.65%, Dice Similarity Coefficient (DSC) 89.74 ± 8.84% and Jaccard Similarity Coefficient 82.39 ± 12.62%. For classification, the accuracy, sensitivity and specificity of diseased lung recognition was 100%. Classification between emphysema and fibrosis resulted in an accuracy of 93%, whilst the sensitivity was 94% and specificity 88%. In addition to the automatic classification of lungs, deviation maps created by the CAD scheme provide a visual aid for medical experts to further assess the severity of pulmonary disease in the lung, and highlights regions affected.