Pathology is medicine's ultimate diagnostic discipline, tasked with separating the myriad benign conditions from the malignancy that kills. But carrying out this task is no simple feat, as there are more than 80,000 diseases identifiable via specimen (urine, blood, and tissue, for example). Currently, pathologists rely on themselves and their colleagues to achieve diagnoses, but with the number of biopsies increasing as the population ages, their capacity is limited. There is a clear need for a streamlined diagnostic process and for computer assistance.
The advent of robotic slide scanners, which convert glass slides to digital, has been a critical step toward the development of whole slide imaging (WSI), the process by which digital slides are not only generated but stored, viewed, and transmitted. One highly successful application of WSI is automated computerized screening for Pap smears1 and the detection of dysplasia in Barrett's esophagus.2 Clearly, computers will not take pathologists' place, but in the clinical lab of the future, a synergy between the pathologist and the computer, with its image analysis capabilities, will lead to a more efficient screening and diagnostic process. There are challenges to achieving this aim: for example, how to store the many terabytes of data, as digital slides tend to be several gigabytes each; how to transmit this information over the Internet; and how to search these large databases. Our focus is on the pathologist's interface to visualize these images.
Viewing digital slides requires specialized software, and manufacturers of different scanning systems have their own customized versions. Most of these are based on map navigation interfaces. These allow the user to zoom (up to a pre-determined level) and pan the image, but not get lost due to a ‘thumbnail’—a low-resolution view of the entire slide—which sits in a corner and marks the user's location. In most viewers, navigation can be carried out either on the thumbnail or directly in the main viewer. The latter, however, is inefficient because at high magnification—40×, for example—a digital slide is the size of a football field, and dozens of mouse clicks would be necessary to go from one end to the other. Different vendors may also include additional side viewers, which contain patient data—such as the patient's history and previous slides—and/or the pathologist's case list for that day. In general, the main viewer, where the digital slide is displayed, can range from the commonly used over-the-counter color monitor (with a resolution of 1280 × 1024 pixels) to a very rare medical grade 6MP color liquid crystal display (3280 × 2048 pixels). Most pathologists believe they can review digital slides using the color display they already have in their offices, but our research shows that this may not be true.
Due to the digitization process, a conventional 1280 × 1024 pixel display shows only about 18% of the field of a conventional microscope when focusing on the same slide at the same magnification.3 At first glance, this result suggests that reading digital slides will intrinsically take longer, as a significantly smaller area is covered at each viewing. Other studies found that it can take up to 60% longer to read cases on the virtual microscope.4 However, if display resolution is the limiting issue, perhaps adding more pixels would alleviate the problem, as larger displays may lead to fields of view that are comparable between the light microscope and the virtual microscope. To investigate this hypothesis, we have developed the virtual reality microscope (VRM), an 11MP display comprising three 27 inch screens3 (see Figure 1).
Figure 1. Virtual reality microscope, shown at the Leeds Institute of Molecular Medicine at St James' University Hospital, UK. A digital slide is shown across three screens, and pathologists can zoom in and out, as well as pan the image as they wish. A very large thumbnail is shown in the right screen, which can be used for navigation.
Analysis of experts' and trainees' performance using the VRM suggested that the large number of pixels eliminated the differences in reading times between the conventional and virtual microscopes. Nonetheless, there were two major errors—with significant impact on patient care—and six minor errors made on the VRM, whereas no errors were made on the conventional microscope.
The occurrence of these errors highlights an important point: that WSI is not the ‘digital analog’ of light microscopy. Rather, it is a completely new way to diagnose disease. It comes with many advantages over light microscopy, such as the ability to store slides with no loss of information, and expedited teleconsultation. But it also has challenges, not least being how best to visualize the slides. Our work suggests that larger displays lead to comparable reading times, and hence to less impact on workflow. However, we also found that diagnostic criteria used in conventional microscopy may not work when translated to WSI. For example, pathologists sometimes did not see very small findings, such as eosinophils (a form of white blood cell), when using a virtual microscope, potentially compromising disease diagnosis. We also determined that pathologists need twice the magnification used in the light microscope to visualize the same details in the virtual system.3 Therefore, our future work will focus on identifying criteria that may lead to diagnostic errors, and on determining the adaptations pathologists need to address these.
This work was partly supported by a grant from the Agency for Healthcare Research and Quality K01 HS018365-01.
University of Pittsburgh
Claudia Mello-Thoms is an assistant professor of biomedical informatics and radiology. Her goals are to understand and model medical image interpretation. Her research interests include, but are not limited to, image perception/interpretation, visual search, whole slide imaging, intelligent tutoring systems, and human-computer interaction.
1. PRISMATIC Project Management Team, Assessment of automatic primary screening on PAPNET of cervical smears in the PRISMATIC trial, Lancet 353, p. 1381-1385, 1999.
2. J. P. Baak, F. J. ten Cate, G. J. Offerhaus, J. J. van Lanschott, G. A. Meijer, Routine morphometrical analysis can improve reproducibility of dysplasia grade in Barrett's oesophagus surveillance biopsies, J. Clin. Pathol. 55, p. 910-916, 2002.
3. R. Randell, R. A. Ruddle, C. Mello-Thoms, R. G. Thomas, P. Quirke, D. Treanor, Virtual reality microscope versus conventional microscope regarding time to diagnosis: an experimental study, Histopathology 62(1), p. 351-358, 2013.
4. D. Treanor, P. Quirke, The virtual slide and conventional microscope--a direct comparison of their diagnostic efficiency, J. Pathol. 213, p. 7a, 2007.