ROC and LROC analyses of the effects of lesion contrast, size, and signal-to-noise ratio on detectability in PET images

Image quality in PET is typically assessed using measures such as contrast recovery, noise variation, and signal-to-noise ratio (SNR). However, these criteria do not directly reflect performance in the clinical use of the ima ges. Lesion detection is a critical task in the clinical interpretation of many PET studies. A receiver operating characteristic (ROC) study is an acc epted method for quantitatively evaluating detection performance with respe ct to factors that influence image quality. ROC and localization ROC (LROC) analyses were conducted to investigate the effects of lesion contrast, SNR , and size on detectability of hot lesions in PET images. Methods: A thorax phantom was imaged with spheres of 3 sizes simulating lesions (0.45, 1.0, and 1.9 mt). The relative activity in the lesions and the total number of c ounts acquired were each varied by factors of 2 to ascertain the effects of contrast and SNR, respectively. Measured attenuation correction and a stan dard reconstruction protocol were used. Three nuclear medicine physicians a nd 6 medical physicists participated as readers, rating each image and indi cating the suspected lesion location. The area under the calculated ROC and LROC curves (A(z) and A(z,LROC)) were used as measures of detection perfor mance. Results: Detection performance was shown to increase from virtually random (A(z) similar to 0.5, A(z,LROC) similar to 0.2) to superior (A(z) > 0.9, A(z,LROC) > 0.9) as lesion contrast was increased by 50% and as lesion SNR was doubled. Detection performance was not seen to vary when compariso n was made using image-based measures alone. Conclusion: This study quantit atively shows that moderate increases in the image-based measures of lesion contrast and SNR give a relatively large increase in the task-based measur e of lesion detection as measured by ROC and LROC analyses. Thus, technique s that give modest increases in lesion contrast or SNR are expected to impr ove detection. Results will be useful in evaluating improvement in detectio n for various reconstruction, acquisition, and data analysis methods that e nhance contrast or noise performance.