Performance Evaluation of Lung Segmentation

Abstract

Segmentation of structures in medical images is challenging due to several factors which include anatomical differences, abnormalities in lung tissue, image noise, and differences in acquisition parameters. Segmentation is one of the key initial components of Computer Aided Diagnosis (CAD) system. Radiologists deal with a heavy workload of having to examine a large number of high resolution computed tomography (HRCT) images. CAD-based systems can lighten their load. and also aiding them as a tool in their diagnostic evaluations. The development of CAD and hence the automatic segmentation is aggressively pursued by researchers worldwide. Therefore it is important to determine the performance or the quality of the automated segmentation developed. Here in this chapter, different segmentation performance evaluation methods for medical images are presented. For most segmentation evaluations, it is important to have comparison done with the gold standard. The gold standard for segmentation involving medical images is the delineation or manual tracing of the region of interest done by a human expert who is preferably a radiologist. The smaller the deviation of the segmentation compared to the human expert, the higher the performance or the quality of the segmentation. The performance evaluation for the segmentation is divided into quantitative and qualitative methods. Most quantitative methods fall into two categories which are area based evaluation methods where the difference between the areas of segmentation and the gold standard are compared, and surface evaluation type where the method evaluates based on the difference between contours of the segmentation and the gold standard. This chapter also discussed a performance evaluation of an automated lung segmentation systems (ALSS) developed by UTM Razak School. This chapter shows the vast performance measures available for determining the segmentation quality. More than one type of performance measure should be used to give a broader and unbiased view of the segmentation quality.