In-vivo detectability index: Development and validation of an automated methodology
The purpose of this study was to develop and validate a method to estimate patient-specific detectability indices directly from patients' CT images (i.e., "in vivo"). The method works by automatically extracting noise (NPS) and resolution (MTF) properties from each patient's CT series based on previously validated techniques. Patient images are thresholded into skin-air interfaces to form edge-spread functions, which are further binned, differentiated, and Fourier transformed to form the MTF. The NPS is likewise estimated from uniform areas of the image. These are combined with assumed task functions (reference function: 10 mm disk lesion with contrast of -15 HU) to compute detectability indices for a non-prewhitening matched filter model observer predicting observer performance. The results were compared to those from a previous human detection study on 105 subtle, hypo-attenuating liver lesions, using a two-alternative-forcedchoice (2AFC) method, over 6 dose levels using 16 readers. The in vivo detectability indices estimated for all patient images were compared to binary 2AFC outcomes with a generalized linear mixed-effects statistical model (Probit link function, linear terms only, no interactions, random term for readers). The model showed that the in vivo detectability indices were strongly predictive of 2AFC outcomes (P < 0.05). A linear comparison between the human detection accuracy and model-predicted detection accuracy (for like conditions) resulted in Pearson and Spearman correlations coefficients of 0.86 and 0.87, respectively. These data provide evidence that the in vivo detectability index could potentially be used to automatically estimate and track image quality in a clinical operation.
