Previous studies of the effects of lossy compression on diagnostic accuracy have focused on the detection of structures [5,7,8,15,19,26]. However, measurement tasks also play a crucial role in diagnostic radiology. Measurements on structures such as blood vessels, other organs, and tumors take a central role in the identification of abnormalities and in following therapeutic response. Radiologists routinely measure almost everything they detect. For example, while diagnosing a fractured bone, they might measure the displacement between the two pieces of bone, or when reporting the presence of metastatic lesions in the liver, they might measure the size of the largest one. Often such measurements are not of great importance to clinical decision-making; but in some examples, they are of extreme significance. In vascular surgery, for example, precise measurements taken on angiograms of the distance from an area of stenosis to the nearest bifurcation in the vascular structure are needed to reduce surgical exposure. In the evaluation of aneurysms, size is an important prognostic feature in any presurgical assessment. Precise measurements on images are increasingly important in those areas where 3D
stereotactic localization can lead to less invasive surgical biopsy methods. For example, in mammography, fine needle biopsy techniques require careful distance measurements in order to place the needle correctly. For our study of the effects of compression on measurement accuracy, we chose to look at measurements of aortic aneurysms, one of the most common areas where size measurements radically affect clinical decision making.
Abdominal aortic aneurysms are usually evaluated with ultrasound, and thoracic aortic aneurysms are evaluated by CT or MRI. In the latter case, the aortic diameter is usually measured manually with calipers. If the aorta exceeds 4 cm in diameter, an aneurysm is diagnosed. A larger aneurysm carries a greater risk of rupture, with approximately 10% risk of rupture for aneurysms between 5 and 10 cm in diameter, and about 50% for aneurysms greater than 10 cm . Rupture is invariably fatal, and so when the aorta measures more than about 5 or 6 cm in diameter, operative repair is usually recommended [6,28]. The clinical decision depends not only on the size of the aneurysm but also on the clinical status of the patient (issues of pain and hemodynamic instability). Dilation less than 5 cm in diameter may be followed conservatively by serial MR imaging studies at 6-month intervals. Observing an increase in the aortic diameter of 0.5 cm over the course of a 6-month interval would be indication for surgical repair.
The study described here had as its goal to quantify the effects of lossy compression on measurement accuracy through experiments that follow closely the clinical tasks of radiologists evaluating aortic aneurysms [24,25]. We wished to examine whether compression maintains the information required for accurate measurements, or whether it leads to inaccuracies by blurring edges or distorting structures. If compression at a certain bit rate caused a 0.5-cm error in the aortic measurement, that would have an impact on the clinical decision, and the compression would be unacceptable. Although we focused on the medical problem of thoracic aortic aneurysms as seen on MR scans, the methodology developed in this research is applicable to any medical task requiring the measurement of structures.
The task studied was the measurement of four primary blood vessels in the mediastinum: the ascending aorta, descending aorta, right pulmonary artery (RPA), and superior vena cava (SVC). A set of 9-bit original MR chest images containing aneurysms and normal vessels was compressed to five bit rates between 0.36 and 1.7bpp. Radiologists measured the four vessels on each image.
In our statistical analyses, we set two gold standards, "personal" and "independent." As discussed in the previous chapter, these represent two methods of establishing the correct size of each blood vessel, that is, the underlying diagnostic "truth" of each image. For each of these gold standards, we quantify the accuracy of the measurements at each compression level by taking the percent measurement error for each image, defined to be the difference between a radiologist's measurement and the gold standard, scaled by the gold standard measurement. This error is examined as a function of bit rate by using the i-test and a nonparametric competitor, the Wilcoxon signed rank test.
Was this article helpful?