Computer-Aided Diagnosis in Radiology: Medical Image Analysis for the Scoring of Chest Radiographs in Cystic Fibrosis

Abstract

Cystic fibrosis (CF) is one of the most common life-threatening genetic disorders worldwide and in Australia, causing considerable morbidity and mortality. Multiple organs are affected, including lungs in which a typical pattern of changes occurs, with progressive inflammatory thickening of the airways and destruction of alveoli. Clinicians assess disease severity using a combined score that includes radiological imaging. The Shwachman-Kulczycki scoring system used in Australia classifies chest radiographs (CXRs) into five categories. Determination of the score is still performed entirely by clinicians, and so an automated scoring system is an innovation that would provide an objective measure of the CXR abnormalities. This study proposed a novel computer-aided diagnostic (CAD) framework for fully-automated scoring of paediatric CXRs in CF. A patch-based method for segmentation was implemented and evaluated. PBS of CXRs achieved median overlap of 0.939 using a 70-image reference set and voting between the 13 nearest neighbours. Performance degraded gracefully, even for large reductions in bit depth and reference set size. Texture features were selected and examined for their ability to discriminate disease severity. Tamura features, local binary patterns, grey-level co-occurrence matrix-derived features and Gabor filtering were used, as inputs to linear discriminant analysis and support vector machine (SVM) classification. A spatial-domain band-pass filter was also developed to enhance the sensitivity of the texture analysis. The overall best classification performance, using Tamura features and SVM, was 75% agreement with clinician scores for classification among three levels (mild, moderate, severe), and 90% agreement for mild-versus-severe classification. This level of performance was similar to the level of human interobserver agreement for the dataset.