The importance of image quality: In the eyes of the beholder?

Abstract

It is tempting to consider image quality as being represented by a “pretty picture.” Think of the differences in the art of Jackson Pollack and Claude Monet. Preferences between the works of these two artists can differ widely and are largely intuitive, emotional, subjective, and observational. Similarly, a consideration of medical image quality is largely embedded in a subjective and observational perspective, one that has dominated the clinical practice of radiology. The emphasis on continuous improvement in image quality has yielded tremendous gains for patient management as, for example, in 1) the ability to distinguish benign from malignant lesions and avoid tissue sampling (eg, cystic renal mass characterization, adrenal mass characterization); 2) the ability to guide the most appropriate therapy in acute or chronic disease states; 3) the ability to avoid an exploratory laparotomy for many conditions. These gains have resulted from important insights based on careful observations from the acquisition of images with fine detail and strong image contrast. The dictum “…maximize contrast resolution and minimize partial volume effects…” was a philosophy drilled into my young malleable mind as a radiology resident and MRI fellow during my training years at New York University Medical Center. This perspective has served many radiologists quite well. It has paved the way to create images of sufficient quality so that the standard of care for certain patient conditions has been transformed and positively impacted. I witnessed such impact as a trainee, perhaps best exemplified in the emergence of using CT images as a means for a definitive diagnosis of appendicitis, what had previously been considered a strictly “clinical” diagnosis 1, 2. In the preimaging era, the diagnosis of appendicitis was based on physical examination and laboratory data. However, it was subsequently shown that the use of high-quality imaging lowered the negative appendectomy rate (the percentage of cases in which surgery revealed a normal appendix) without increasing the incidence of abscess or perforation, as could happen with waiting too long to make the diagnosis and sustaining unwanted complications 3. Such meaningful impact led to the development of an MR protocol that achieves the same benefits for pregnant patients and children with abdominal pain, challenging patient populations, for whom avoidance of ionizing radiation is desirable 4-6. The “abdominal pain in pregnancy” MR protocol was not feasible until image quality was sufficient to reliably image the normal and abnormal appendix, as well as other associated structures. This established an exam that allows for an accurate diagnosis of appendicitis and alternative causes for pain while avoiding ionizing radiation 7. Motion immune sequences such as HASTE and SSFSE with thin sections were critical in eliminating bowel motion, providing strong contrast and generating images where the wall thickness of the appendix could be measured. Image contrast has been augmented through the use of diffusion-weighted imaging (DWI) for further improvements in diagnostic accuracy 6. The example of appendicitis shows how critical image quality is to improving patient care and well-being. Wait a minute…. Could I have just implied that DWI of the appendix might represent an improvement in image quality? As an interesting exercise, compare the two provided images and decide which has the better image quality (Fig. 1). While I will maintain that there is a critical need for continuous improvement of imaging quality, the previous example exposes the controversy in defining image quality. A simple description of image quality is a subjective notion based on an aesthetic appreciation or preference. Such a judgment is difficult to categorize, refine, and reproduce but has been a mainstay of observational clinical radiology. While examples on the extremes are often “obvious,” those comparisons with subtle distinctions can be challenging and require detailed information (Fig. 2). Such an example helps expose the vagaries and quirkiness that thwart attempts to establish utility in the subjective domain. When producing an image, the goal is to make the most accurate representation of reality and, ideally, to facilitate the conspicuity of an abnormality or variant. In this context, a high quality image would have minimal or no artifact, high spatial resolution, and strong image contrast relevant to the disorder being considered, when that is known or suspected in advance. In reality, the condition prompting an evaluation is often unknown at the time that an exam is requested. Furthermore, the condition may be associated with a variety of disorders or associated alterations. Therefore, in clinical practice an imaging strategy needs to be sufficiently broad to project a wide range of imaging features capable of capturing the altered state of reality. Complicating this is how the demands on time and efficiency influence the inevitable trade-offs among the ideal attributes that we seek in an image. In an effort to achieve aesthetic differentiation or a perception of preference in MR imaging, a variety of strategies are employed by manufacturers, radiologists, and imaging scientists. Many strategies provided by manufacturers are concealed, veiled as “trade secrets” resulting in what I term “soft differentiation.” Others are well intentioned to boost imaging efficiency. An example of the latter can be found in pulse profile changes that are automatically adjusted to accommodate a desired resolution, acquisition time, and comply with limits on specific absorption rate (SAR). In fact, such strategies may result in images that have relative appeal to an observer or equipment purchaser but may actually obscure detail. One example is an unintended degradation in the point spread function from an automated SAR reduction pulse sequence modification that could compromise spatial resolution and render a critical detail difficult or impossible to detect and characterize. Postprocessing options that impact the aesthetics of the images, from subtle to substantial, have been available. One vendor's implementation resulted in a series of options that imparted a range of smoothed appearances to the images. In one of my former practices, different radiologists selected different algorithms to be applied to the images of the cases scheduled for their interpretation! To the best of my knowledge, there was never a consensus on the optimal smoothing algorithm since individual preference was the determinant (not outcome or some other more meaningful endpoint). Aesthetic preference as a basis for image quality lacks the robustness and value needed to serve as a meaningful determinant. In clinical interpretation studies, reader preferences for certain image features can provide insights into the potential utility of a new imaging strategy: features such as the edge clarity of relevant organs, artifact severity, detection of anatomy or pathology, and uniformity of fat suppression, are examples 8. However, as complexities in image acquisition strategies emerge, such as those to reduce scan time, the results may be less straightforward. For example, in a recent study evaluating subjective image quality, the introduction of compressed sensing to reduce acquisition time was shown to accelerate certain neuroimaging sequences without severe loss of clinically relevant information 9. However, in that same study, for those sequences with coarser spatial resolution and/or higher acceleration, artifacts degraded the quality of the reconstructed image to a point where they were felt to be “…of little to no clinical value” 9. Such disparate results reflect the challenges present in precisely defining “image quality” and the implications for patient care or extraction of useful data for research purposes. The appearance of the MR image is the result of necessary tradeoffs between speed, spatial resolution, signal-to-noise ratio (SNR), and image artifacts. The semiquantitative terms SNR and contrast-to-noise ratio (CNR) are often used for comparative insights. For example, the SNR benefit for quantifying vessel cross-sections has been demonstrated 10. While SNR and CNR evaluations add rigor, they remain in the technical domain of image quality and do not offer an index of efficacy with respect to patient management or data extraction. Furthermore, the technical frame of reference has become far more complex to analyze with the introduction and widespread use of parallel imaging, for which a variety of reconstruction algorithms yield different noise distributions and artifacts 11, 12. Another concept of image quality is defined by the ease with which image features, critical to making a management decision, can be extracted and distinguished from a set of background features. This extraction can occur as a result of human interaction, automated machine interaction, or combinations of the two. In this conceptualization, quantitative imaging strategies may benefit from high image quality, a condition that can be met when that set of characteristics facilitates the extraction of information to render a specific diagnosis or, better yet, a cost-effective management decision. Indeed, quantitative analyses may allow for the extraction of features that are difficult to appreciate by a large number of radiologists, if not impossible to detect by human perceptive skills alone. In this sense there can be an enhancement of the radiologists' role in generating a more decisive opinion regarding diagnosis or management. The opportunity to automate the extraction of key imaging features is increasing and depends on the ability to capture and differentiate key determinants within an image. As an example, the ability to automatically derive prostate gland volumes has important implications to enhance management options for both benign and malignant disorders 13. “Automated” extraction, however, often requires an operator to set the conditions allowing for the computer processes to then proceed. And so, for the vast majority of tasks related to imaging, human perception is a necessary tool. An appreciation of that reality has prompted research effort to incorporate aspects of human visual response, sensitivity, and characteristics in computing quality indices 14. Regardless of how we define image quality, radiologists generally know a good image when we see it, and many radiologists find the ability to generate such an image rewarding. But it pays to keep in mind that an “exquisite” image may 1) leave the interpreter without the ability to make a helpful diagnosis or impact the management plan [insufficient relevant detail]; 2) facilitate an elegant diagnosis but one that may have no effective treatment nor measurable impact on outcome (eg, some brain neoplasms) [sufficient relevant detail but no practical value]. Therefore, a high-quality, exquisite image may, at times, better satisfy the radiologist rather than the referring physician or patient; in this context, striving for the best image possible may be a misdirected effort. With extended needs and opportunities to probe biologic processes and demonstrate disorders at the earliest possible timepoint (smallest size?), the demands of the information yield from the imaging endeavor broaden. Does this then portend an asymptote to the yield of image quality as it is generally considered? A different approach looks at the impact of imaging on the patient management process 15. Using a hierarchical model of efficacy, Fryback and Thornbury 15 presented an organizing structure for appraisal of the literature on efficacy of imaging (Fig. 3). Demonstration of efficacy at each lower level in this hierarchy is logically necessary, but not sufficient, to assure efficacy at higher levels. Level 1 concerns technical quality of the images; Level 2 addresses diagnostic accuracy, sensitivity, and specificity associated with interpretation of the images. Next, Level 3 focuses on whether the information produces change in the referring physician's diagnostic thinking. Such a change is a logical prerequisite for Level 4 efficacy, which emphasizes the effect on the patient management plan. Level 5 efficacy studies measure (or compute) the effect of the information on patient outcomes. Finally, at Level 6, analyses examine societal costs and benefits of a diagnostic imaging technology. Most assessments of image quality dwell in Levels 1 and 2. However, movement into the higher levels of efficacy deemphasizes the traditional notions of image quality. In their classification, Level 6 analyses conceptualize efficaciousness to the extent that it is an efficient use of societal resources to provide medical benefits to society 15. This is when considerations of cost are first encountered in that rubric. It can be argued that the impact of cost can be considered earlier in the model, particularly in our current reimbursement climate with ever-increasing costs being shifting to the patient. An earlier consideration of cost is relevant and often of greater immediate concern for evaluating the utility of imaging in a research paradigm as well. Thus, image quality can be represented by an index of content value/cost. In a clinical context, content value can be determined by the accuracy with which an interpretation renders appropriate management, including the exclusion of relevant alternative considerations; cost can be determined with total cost accounting, including all associated capital and human resources, as well as time considerations. In the clinical sphere, time, at a minimum, involves the image acquisition, the movement of the patient through the system (including the necessity to repeat an exam or impact additional studies), the ease with which the radiologist can formulate a diagnosis or management plan, and the communication of results. In the clinical research sphere, image quality can also be considered as content value/cost. Here, content value can be determined by the ease with which salient observations can be made and reproduced and the time needed for acquisition, evaluation, and information extraction. In this editorial I have acknowledged and discussed the challenges in defining image quality, shown examples of how image quality can make meaningful differences, and offered alternatives to consider when judging image quality. These efforts lead me to a simple conclusion: in our present state of MRI technology and how we interact with it, image quality matters greatly, and is crucial for the efficacious use of image data. Since the importance of image quality rests in the value it delivers, our goal is to transfer this from the current state, in the eyes of the beholder, to a future state in which it is readily apparent to most, by virtue of its reproducibility and widespread efficacy. Now there's something we'll know when we see it! Neil M. Rofsky, MD Department of Radiology Advanced Imaging Research Center UT Southwestern Medical Center Dallas, TX, USA