Automatic segmentation in CMR - Development and validation of algorithms for left ventricular function, myocardium at risk and myocardial infarction

In this thesis four new algorithms are presented for automatic segmentation in cardiovascular magnetic resonance (CMR); automatic segmentation of the left ventricle, myocardial infarction, and myocardium at risk in two different image types. All four algorithms were implemented in freely available software for image analysis and were validated against reference delineations with a low bias and high regional agreement. CMR is the most accurate and reproducible method for assessment of left ventricular mass and volumes and reference standard for assessment of myocardial infarction. CMR is also validated against single photon emission computed tomography (SPECT) for assessment of myocardium at risk up to one week after acute myocardial infarction. However, the clinical standard for quantification of left ventricular mass and volumes is manual delineation which has been shown to have a large bias between observers from different sites and for myocardium at risk and myocardial infarction there is no clinical standard due to varying results shown for the previously suggested threshold methods. The new automatic algorithms were all based on intensity classification by Expectation Maximization (EM) and incorporation of a priori information specific for each application. Validation was performed in large cohorts of patients with regards to bias in clinical parameters and regional agreement as Dice Similarity Coefficient (DSC). Further, images with reference delineation of the left ventricle were made available for future benchmarking of left ventricular segmentation, and the new automatic algorithms for segmentation of myocardium at risk and myocardial infarction were directly compared to the previously suggested intensity threshold methods. Combining intensity classification by EM with a priori information as in the new automatic algorithms was shown superior to previous methods and specifically to the previously suggested threshold methods for myocardium at risk and myocardial infarction. Added value of using a priori information and intensity correction was shown significant measured by DSC even though not significant for bias. For the previously suggested methods of infarct quantification a poorer result was found in the new multi-center, multi-vendor patient data than in the original validation in animal studies or single center patient studies. Thus, the results in this thesis also show the importance ofusing both bias and DSC for validation and performing validation in images of representative quality as in multi-center, multi-vendor patient studies