In this chapter and the following two chapters, results are presented for three data sets: computerized tomography (CT), magnetic resonance (MR), and mammographic images. As will be seen later, these three studies provide examples of the detection, localization, measurement, and management aspects of a radiologist's interpretative functions.
The CT study involved two different sets of chest images. In one, the diagnostic task was the detection of abnormally enlarged lymph nodes, and in the other, the task was to detect lung nodules. Thirty patient studies were used for each task. The CT images were compressed using pruned predictive vector quantization  applied to 2 x 2 pixel blocks . This method involves no transform of the data. Vector quantizers are often designed for a training set of representative images that can provide information about the statistics such as the spatial correlations that are typically found in those images. In such a situation, the compression algorithm will perform best for images that are similar to those used in the training set. For this study twenty CT images of the mediastinum were used in the training set for detecting enlarged lymph nodes, and 20 CT lung images were used in the training set for detecting lung nodules. All 512 x 512 pixel images were obtained using a GE 9800 scanner (120 kV, 140 mA, scan time 2 seconds per slice, bore size 38 cm, field-of-view 32-34 cm). Although no formal research was undertaken to determine accurately what constitutes "representative" CT images, two radiologists were consulted concerning the typical range of appearance of adenopathy and nodules that occurs in daily clinical practice. The training and test images were chosen to be approximately representative of this range, and included images of both normal and abnormal chests. The lung nodules ranged in size from 0.4 to 3.0 cm, with almost all nodules between 0.4 and 1.5 cm, and the abnormal lymph nodes were between 0.6 and 3.5 cm. The study also had a lower percentage of normal chest images than would be encountered in daily practice.
For each study (lymph nodes, lung nodules), the original 30 test images were encoded at six compression levels: 0.57, 1.18, 1.33, 1.79, 2.19, and 2.63 bits per pixel (bpp). The original test images are considered to be 11-bit data. Figure 2 shows an original 11-bpp CT lung image to which the "windows and levels" contrast adjustment has been applied. Although the scanner was capable of producing 12-bit data, it was found for this data set that the 12th bit was never used. Patient studies represented in the training set were not used as test images, and the results reported on SNR, subjective quality, and diagnostic accuracy are based only on test images.
3.2 MR Study
In the MR study, the diagnostic task was to measure the size of blood vessels in MR chest scans, as would be done in evaluating
aortic aneurysms. The study had as its goal to quantify the effects of lossy compression on the accuracy of these measurements [46,47]. As in the CT study, the image compression scheme was predictive pruned tree-structured vector quantization, although in this case it was applied to blocks of 2 x 4 pixels.
The training data of 20 MR chest scans were chosen to include a wide range of both aneurysms and normal vessel structures. An additional 30 scans were chosen as test images. All images were obtained using a 1.5-T whole body imager (Signa, GE Medical Systems, Milwaukee, WI), a body coil, and an axial cardiac-gated Tl-weighted spin echo pulse sequence with the following parameters: cardiac gating with repetition time (TR) of 1 R-R interval, echo time (TE) of 15-20 msec, respiratory compensation, number of repetition (NEX) of 2, 256 x 192 matrix, slice thickness of 7 mm with a 3-mm interslice gap.
The compression rates for this study were 0.36, 0.55, 0.82, 1.14, and 1.70 bpp on the 30 test images. These bit rates are represented by compression levels 1-5. The original scans at 9.0 bpp are represented by level 6.
Figure 3a shows an original 9.0 bpp MR chest scan. Figure 3b shows the same image compressed to 1.14bpp, and Fig. 3c shows the image compressed to 0.36 bpp.
Was this article helpful?