Alisa I. Walz-Flannigan, PhD, Mayo Clinic; Kenneth A. Fetterly, PhD; Steve G. Langer, PhD; Beth A. Schueler, PhD

Bit depth reduction sometimes occurs in the archiving and display of medical images and can have an impact on image quality. The extent of image degradation for a given degree of bit reduction can be assessed by measuring the associated quantization noise and comparing it with the expectedly dominant quantum noise in digital radiographic images. This tool can be used as to establish acceptable levels of bit reduction for a given modality and clinical image type.

In our clinical practice, we are operating many imaging systems whose native output is greater than 12 bits. In principle, by windowing and leveling these images, radiologists are able to access the full 12+ bits in an image, even if the display is limited to 8 bits. However, when these images are put into our PACS system, they are subsequently scaled and truncated to 12 bits when displayed. We wished to assess the impact of this bit reduction on the image quality for specific clinical exam types in a digital radiographic (DR) system. We asked the question, is there useful signal in those additional bits? And, if not, to what degree could we safely bit truncate a given image without significant loss of information?

Bit-depth reduction in an image is equivalent to increasing the quantization of gray levels. This quantization always introduces error (noise) and the degree of noise introduced is proportional to 0.3 times the distance between quantization levels.[1] Whether this will have an impact on image quality depends on its magnitude, relative to the other noise sources in the image. If the quantization noise introduced from a bit-depth reduction is significantly less than quantum noise in a DR system, it’s of negligible impact. This is the same as saying that if the information contained in the additional bits is not of greater magnitude than quantum noise in the image, it is not visible. We wanted to see if this was the case for our natively 14-bit chest DR images, and how far the images could be safely bit reduced before impacting image quality.

This work will be extended to include other exam types and modalities with outputs greater than 12 bits as used in our clinical practice.

Images of an anthropomorphic chest phantom were acquired with a typical clinical technique on a GE Definium 8000 DR unit [GE Healthcare, Milwaukee, WI]. Because we wanted to assess the impact of bit reduction on clinical images, the following procedure was done with “for presentation” images (in contrast to “raw” or “for processing” images) with our chest specific image processing already applied. Two identically acquired images were subtracted from each other and used to measure the quantum noise for relatively uniform regions of interest (ROIs) in the lung and mediastinum. The quantum noise is measured for an ROI by dividing the standard deviation in the subtracted-image ROI by the square root of 2.

Quantization noise from bit reduction was measured by using one of the original 14-bit images to create simulated bit-reduced images (bit depths between 7 and 13-bits). The bit-reduced images were then rescaled to the same pixel value range as the original image and subtracted from the original image. This difference image was used to measure quantization noise in the same regions as were used to measure the quantum noise.

The total noise in the image is the quadrature sum of the noise from different sources. Assuming the presence of only quantum and quantization noise, we calculated the total noise in the lung and mediastinum ROIs. To better understand the effect of the additional noise on the image, we calculated the signal-to-noise ratio (SNR) for the lung and mediastinum, as well as the percent change in SNR resulting from the additional quantization noise.

While observer studies are needed to establish what tolerances are acceptable for magnitude of change in SNR for a given task, our technique provides a metric for assessing the physical impact of bit reduction on the image. If one assumes that a change in SNR of less than 1% will be imperceptible to the observer, then converting our 14-bit chest DR images to 12 bits will not have an impact on the practice. We can use this technique to better understand the physical impact of bit-depth reduction on image information content across different modalities and image types.

1. Jahne B. “Digital Image Processing.” Berlin: Springer-Verlag; 2005.