Assessment of image quality and detectability in computed tomography: Inter-laboratory comparison of Fourier-based...

The task-based assessment of image quality in computed tomography (CT) is increasingly used for benchmarking, optimizing acquisition and reconstruction parameters, and predicting observers' performance. To achieve consistent and reproducible results rely on standardized methodologies and guidelines.

This study performed an inter-laboratory comparison to evaluate the consistency of CT image quality metrics and assess inter-laboratory variability.

Six series of images of a Catphan 500 phantom were acquired at three dose levels (1.5/5.0/15.0 mGy) with two (standard and high-resolution) kernels and sent to seven independent laboratories. Participants calculated the task transfer function (TTF) for spatial resolution and noise power spectrum (NPS) for noise magnitude and texture using standardized Fourier-based analysis on the shared images. The TTFs and NPS were used in the non-prewhitening observer model without (NPW) and with eye filter (NPWE) to calculate a detectability index for small (2 mm) and large targets (10 mm).

The difference observed in TTFs resulted mainly from the different noise reduction techniques used during the calculation. The frequency corresponding to the TTF value 50% varied from 0.27 to 0.43 mm^-1 for the standard images and from 0.73 to 0.92 mm^-1 for the high-resolution images. Variability in NPS primarily stemmed from the choice of regions of interest location in non-stationary noise images that lead to magnitude differences up to 40% for the standard kernel and 70% for the high-resolution kernel. The combined differences in TTF and NPS resulted in differences up to 39% in the NPW detectability index. The large variability in the human visual transfer function and visualization parameters increased the differences by a factor of 6.25 in the NPWE model and highlighted the lack of standardization.

Adapting TTF and NPS to a more complex and localized analysis of CT images with non-uniform CNRs, rather than a simple global evaluation, increases computational variability and requires a move towards more standardization of metrics. In the absence of complete standardization of image quality measurement, documentation of methods and all calculation parameters is essential for more robust evaluation of CT image quality.