Portal de Periódicos da CAPES

Practice Evolution: Decentralized Computer-Assisted Immunohistochemical Image Analysis

2009; American Medical Association; Volume: 133; Issue: 4 Linguagem: Inglês

10.5858/133.4.597

1543-2165

Richard Friedberg, Liron Pantanowitz,

Radiomics and Machine Learning in Medical Imaging

In 2007, the Futurescape of Anatomic Pathology meeting provided plenty of discussion about the technologic future of the field. Toward the end of that meeting, a “Reality Check” discussion provided a critique of the meeting with respect to the day-to-day activities of practicing pathologists.1 At this year's conference, we are providing an update on our institution's attempts to apply some of the tools and techniques presented here last year. Some tools and techniques have worked, and some have not. The introduction of new technology often involves a steep learning curve. Because these tools and technologies are not simply “plug-and-play,” pathologists, technologists, and administrators must recognize that we cannot simply adopt and implement them. We must validate, verify, and make operational these technologies, which are indeed works-in-progress.The internal environment at Baystate Health (Springfield, Massachusetts) is well suited to implementing new technologies. We are a subspecialty-focused and academic-private practice and hybrid model, including extensive anatomic, clinical, and molecular pathology sections integrated into 1 department. Annually, the department processes 40 000 to 50 000 surgical specimens and 6 to 7 million laboratory tests. We also have an average of 1 new breast cancer case per day. Thus, we have ample raw material to test the emerging tools and technologies.Our external environment is much the same as it is for all of pathology: increased technologic innovation coupled with increased biologic information in a background of increased demands from patients, payors, and providers. All of us are dealing with more and more information to ingest, digest, and synthesize. From a high-level perspective, these increased clinical demands are leading to a convergence of 2 independent, long-term trends in anatomic pathology; these trends are evident in the application of computer-aided image analysis (CAIA) in cases of breast cancer.The first, relevant, long-term key trend for anatomic pathology is its evolution along clinical pathology lines. Specifically, anatomic pathology is moving away from its traditional teaching, which indicates that a particular image is diagnostic of a particular disease primarily because the viewing pathologist was once instructed by a more eminent pathologist that similar images were diagnostic of that condition. This mode of knowledge transfer is akin to a guild model in which an entry-level artisan is apprenticed to and taught by a more senior artisan who is deemed an expert in the art of the craft. The fundamental techniques based on hematoxylin, eosin, paraffin, and formaldehyde have undergone minimal changes during the past century. In recent decades, however, progress has been made, and pathologists are using newer stains and tools, such as immunohistochemistry (IHC) and fluorescent in situ hybridization, to analyze formalin-fixed, paraffin-embedded tissue. Recent trends are clearly headed toward a more quantitative, reproducible, validated, specific, and reliable approach.Image analysis in anatomic pathology has already started shifting from qualitative (eg, positive or negative), to semiquantitative (eg, 0+, 1+, 2+, 3+), to even more quantitative in some areas (eg, copies per cell). Clinical pathology underwent this same transition many years ago. Curiously, a brief review of the history of IHC and enzyme-linked immunosorbent assays shows that both started around the same time as methods of using an antibody to identify an antigen. With IHC, the antigen expression was identified by using an antibody-linked enzyme, such as alkaline phosphatase, to generate color at the site of the bound antigen. Under the microscope, the presence or absence of the generated color was assessed by an anatomic pathologist to determine whether the stain was positive or negative at the location in question. Determining the adequacy of the “darkness” (ie, immunoreactivity) of the stain and the location within the architecture of the cell or tissue became fundamental skills of the pathologist. For years, this qualitative approach was deemed diagnostically sufficient. In time, the strength of the staining became clinically significant, and therefore, anatomic pathologists typically began to grade staining intensity on a scale from 0+ to 3+. As in the classic example of HER2/neu, such semiquantitative assessments may be inadequate and subject to marked interobserver and interassay variability.The same technology underlies both IHC and enzyme-linked immunosorbent assays, which took hold in clinical pathology. In part, because clinical pathology specimens are typically fluids (versus tissue in anatomic pathology), antibody-antigen complexes can be readily counted, and thus, enzyme-linked immunosorbent assays have quickly become quantitative. Effectively, the basic technology that yielded a stain in anatomic pathology has resulted in an assay in clinical pathology. As anatomic pathologists see how much their practice has changed during the past quarter century due to IHC, so too have clinical pathologists witnessed such changes due to enzyme-linked immunosorbent assays. Ironically, in starting to quantify HER2/neu, we are attempting to convert a technology initially established as a qualitative tool in anatomic pathology and are now recasting it to be quantitative. We are seeing stains becoming assays and qualitative interpretations becoming quantitative assessments. Moreover, we are seeing results directly tied to treatment and not just to prognosis, and as previously mentioned, we are moving away from the guild mentality with “anointed” experts.The trend of anatomic pathology's development along the lines of clinical pathology is also evident in the development of personalized medicine. Because pharmacologic treatment is increasingly dependent on the extent of expression for a particular biomarker in the setting of a specific disease, both the exact diagnosis and accurate quantification of the biomarker have become more important. More than ever, a pathologic diagnosis needs to be accurate, reliable, precise, and reproducible. The prototype for this example is breast cancer and HER2/neu overexpression for treatment with Herceptin (trastuzumab, Genentech, South San Francisco, California). The only reason to treat with Herceptin is if the patient's breast cancer overexpresses the biomarker HER2/neu (the drug target). Pathologists may debate internally whether the optimal methodology to assess HER2/neu expression is to count the number of copies of the c-erbB-2 gene in the DNA of the tumor or to estimate the density of the HER2/neu protein on the cell surface. However, those are questions of methodology, not questions of right versus wrong. In this example of personalized medicine, the anatomic pathologist uses new applications as laboratory tools to predict and determine treatment. This scenario is actually more like the traditional clinical pathology perspective.The second key trend in the practice of anatomic pathology is evolution along radiology/imaging lines. Both fields were established using analog images. Until the mid-1990s or so, the typical procedure for obtaining a radiology image was as follows: (1) position the patient between an x-ray source and a piece of x-ray sensitive film; (2) pass short wavelength photons (x-rays) through the area of interest to expose the film (silver emulsion on a flexible backing); (3) develop the film in a series of chemical baths; and (4) deliver the black and white image to a radiologist for reading on a light box. Similarly, the typical procedure for obtaining an anatomical pathology image was as follows: (1) obtain a specimen from the patient; (2) process the tissue through a series of chemical baths to create a formalin-fixed, paraffin-embedded tissue block; (3) place thin slices of the tissue on glass slides and stain them; (4) position these stained slides between a light source and a series of magnifying lenses (microscopy); and (5) pass visible wavelength photons (white light) directly through the area of interest for reading by a pathologist. In both cases, an analog image was created and was ultimately interpreted by a physician.Recent market and technology forces strongly trend away from analog imaging and toward digital imaging. Initially, similar to using a fax machine, the analog images were scanned to create transportable digital images, promising diminished physical storage requirements. Subsequently, digitally acquired images eliminated the need for film and the chemicals used to develop the film. The film was replaced with sensors to detect the short wavelength photons, and processors converted the signal to an image that radiologists and other physicians were comfortable interpreting. Radiology has adopted direct digital image acquisition. This change remains the fundamental difference between imaging in anatomic pathology and imaging in radiology. With the new technology, because the image is electronically created and acquired, a final image with diagnostic quality can be instantly available on a monitor in the immediate examination room or remotely in Idaho, India, or Australia. Telediagnostic capabilities (teleradiology), an inadvertent result of the digital revolution, have irreversibly changed radiology at its core. In fact, a large part of the ongoing technologic evolution of imaging in radiology has been in the handling, processing, transmission, display, and storage of digital images.The standard digital imaging and display technology used for imaging in radiology/imaging today has little counterpart in pathology. Pathologists still work with photons passing through a stained piece of tissue. Certainly, pathology images can be digitally recorded with a camera fixed to a microscope. Recently, whole slide imagers that can scan a slide in a few minutes and then generate a digital image have been more available. However, obtaining any of these anatomic pathology images still requires tissue procurement, processing, immersion in chemicals, glass slides, stains, light sources, microscopes, image-acquisition or conversion devices, and, not to mention, hours of preparation time by technologists. Furthermore, all of the aforementioned steps are not presently standardized. Recognizing the fundamental differences between digital and analog acquisition of images is critical to understanding the technologic future of pathology.Only when the digital image is obtained can the image processing and handling wizardry, which are so prevalent in radiology/imaging, begin to have an impact in pathology. That impact has yet to be realized. Clearly, the digitization of images allows for new applications. However, the digitization of imaging in radiology was not accomplished primarily for the purpose of better quality. The digitization of imaging in radiology was sold to hospital administrators as a cost savings measure to reduce the cost of maintaining a 7/24/365 file room that generated significant expenses for labor, developing chemicals, film, and physical storage. Digital radiology tools were justified by reduced expenses. The improved quality and image handling were essentially unintended consequences. Nonetheless, from the perspective of both the patient and the physician, the improvements have been decidedly positive.What does this portend for future technologic development in anatomic pathology? Is pathology going to benefit from the improvements in productivity, quality, and throughput that have resulted from the digital revolution in radiology? There will be new applications in pathology; however, because of the aforementioned fundamental differences in image acquisition and value additions, the adoption path will be different. Nevertheless, the degree of impact will be similar.The value additions generated by digital technologies for imaging in radiology primarily involve image handling and display. Today, most radiology images are handled by picture archive and communication systems, which process, collate, organize, display, and store images. Techniques, such as coregistration and convergence imaging, have added a novel level of complexity by allowing for the accurate overlay of multiple images. Not only can a radiologist view a mass on a computed tomography scan on a monitor, but he or she also can overlay a positron emission tomography scan to determine whether the mass (computed tomography image) is metabolically active (positron emission tomography image). Although for many years pathologists have had the capability to use multiple antibodies or stains on a single slide, the ability to turn on and off specific markers would allow pathologists to determine more easily whether individual cells, as opposed to clumps or masses, coexpress markers, such as HER2/neu, estrogen receptor (ER), and progesterone receptor (PR). No longer would a pathologist need to identify an area of interest on one slide, remove that slide, replace it with another slide (with a different piece of tissue from a different level), and then try to find the same area of interest on the second stained slide.Other techniques, such as digital subtraction, windowing, and image manipulation, allow radiologists to manipulate an image in fascinating ways. For example, they can “remove” bones, alter contrast, and rotate something around an axis. Radiologists can create a dynamic image or “movie” from a series of static images to better appreciate the gross architecture of a mass or defect. Imagine what a pathologist could do with real-time imaging capabilities, such as subtracting specific cell types, stroma, or normal architecture from an image. He or she would have a new way of looking at the same lesion. The potential benefits of dynamic imaging in pathology are intriguing, to say the least.The 2 aforementioned long-term trends are mutually reinforcing. The added attention to standardization, precision, and quantification is strengthened by tools for analysis of digital images. Today, CAIA is most readily evident in ER, PR, and HER2/neu assessments in breast cancer. For prognostic purposes in breast carcinoma, the standard practice is determination of biomarkers, such as ER, PR, and HER2/ neu, by IHC and/or fluorescent in situ hybridization. Currently, in cases of breast cancer, the accepted, traditional gold standard refers to a pathologist or pathologists using manual microscopy to determine an immunohistochemical semiquantitative score for biomarkers. This practice involves a preanalytic phase (ie, tissue preparation as in fixation and processing), an analytic phase (ie, immunohistochemical staining), and a postanalytic phase (ie, quantification and reporting of results). Problems related to each of the aforementioned steps, such as lack of standardization, makes quantitative IHC analysis difficult to perform.2 Of interest, one study revealed that discrepancies between HER2/neu IHC and fluorescent in situ hybridization were most often due to manual interpretation and not due to reagent limitations.3 Interobserver and intraobserver variability in scoring is particularly notable with borderline and weakly stained cases. Moreover, there is added subjectivity and fatigue associated with human scoring. This is troublesome given that accuracy is vital when evaluating breast biomarker status to ensure appropriate therapeutic intervention. The lay press has expressed concerns about inaccuracies in breast biomarker testing, especially when pathologists are “required to make judgment calls.”4 Pathologists are also faced with the potential threat of having to send their cases to outside laboratories for routine breast biomarker studies. Can CAIA be used to provide more accurate and reproducible scoring of IHC? In other words, why not give all pathologists the same yardstick for scoring IHC in breast cancer cases?Guidelines published by the American Society of Clinical Oncology/College of American Pathologists for HER2/neu testing in breast cancer indicate that image analysis can serve as an effective tool for achieving consistent interpretation.5 The guidelines also indicate that a pathologist is required to confirm image analysis results and that image analysis equipment, including optical microscopes, must be calibrated and subjected to regular maintenance. Moreover, CAIA procedures must be properly validated. Similar recommendations for using image analysis systems to enhance reproducibility of scoring were published following the Canadian National Consensus Meeting on HER2/neu testing in breast cancer.6 Several companies reported US Food and Drug Administration clearance for CAIA with respect to breast biomarkers. Early skeptics stated that subjective scoring of slides was simple and rapid and should not be replaced by expensive and time-consuming CAIA in daily practice.7 Early studies showed that CAIA was no better than visual analysis.8 Only a few studies found CAIA and manual scoring to be comparable.9 However, most published data have shown that CAIA is in fact superior to manual methods.21011Prior studies show that CAIA is more effective, consistent, and precise than manual scoring. Investigators stress that the expense of CAIA may be hard to justify when volumes are low. Moreover, image analysis may increase time requirements and frequently requires interactive input from pathologists. Problems encountered with CAIA include discrepancies associated with low-level staining, artifacts (eg, dust), interfering nonspecific staining in selected areas, and erroneous low scores generated by small amounts of stained tissue. To our knowledge only one study has compared different CAIA technologies and reported agreement between these systems.12 CAIA systems rely on algorithms that involve several steps, including the following: color normalization, background extraction, segmentation, classification and feature selection, object-oriented (morphology-based) separation of tissue elements (eg, tumor epithelium from stroma), identification of region of interest (subject to further image analysis), and quantification of results (ie, diagnostic score).Given the aforementioned argument in favor of adopting CAIA for the analysis of breast IHC, we implemented and validated such a system in the surgical pathology department at Baystate Health. Our pathology department, which is composed of distant medical centers, deals with a significant volume of breast cases. Our intention was to mimic our daily practice by decentralizing this process, that is, to avoid centralized image analysis being performed by just one pathologist. We took into consideration our bandwidth limitations, lack of whole slide imager, and professional reluctance among pathologists to read digital images. All pathologist thick-client workstations were upgraded and fitted with a standard digital camera on their microscopes. All digital cameras were calibrated. A key component to the success of this project was the standardization of camera acquisition settings. Computer-aided image analysis software (Pathiam, BioImagene, Inc, Cupertino, California) was installed as a Web-based application along with networked servers. The workflow mirrored our daily surgical pathology practice, namely as follows: (1) for each run, an advanced histotechnologist uploaded into the system control IHC images, which were verified by the director of IHC; (2) participating pathologists imaged selected fields of view for their routinely stained ER, PR, and/or HER-2/neu breast IHC cases; (3) pathologists uploaded and analyzed their cases on their workstations using CAIA software; and (4) they generated reports containing quantitative IHC results. All digital images were acquired at the same magnification, and a uniform JPEG file format was used. Individual images were analyzed with reference to a default control parameter set defined for that IHC run. ER and PR nuclear staining was analyzed using the Allred scoring system (ie, proportion + intensity = total score). HER2/neu membranous staining was evaluated using American Society of Clinical Oncology/College of American Pathologists recommendations (ie, 0+, 1+, 2+, and 3+). Manual scoring by pathologists and CAIA were compared with respect to IHC score and time to perform the analysis. Fluorescent in situ hybridization for HER2/neu was obtained for a subset of cases.In most cases, we found that ER and PR scores were concordant. Rare discordant cases between pathologists and CAIA for ER and PR cases was attributed to CAIA erroneously quantifying nonspecific cytoplasmic staining of some tumor cells (ie, false-positive results). For HER-2/neu IHC results, there was excellent concordance between manual and CAIA scoring. There was also good correlation between manual IHC scores, CAIA IHC scores, and fluorescent in situ hybridization results. Heterogeneously stained cases proved difficult for pathologists to analyze, as they were uncertain which field of view to select for subsequent analysis. Immunohistochemistry analysis using CAIA proved to be far more time consuming than manual scoring. Based on this validation study, we conclude that decentralized CAIA for IHC, which is designed to mimic daily surgical pathology workflow in practice, is feasible. However, it was clear that image acquisition with this model requires standardization. Both analog and digital image acquisition systems (ie, cameras) are subject to drift over time.13 Calibration is required to adjust for several variables, such as the light source, color spectrum, and analog-to-digital conversion process. Our experience further indicates that pathologists must supervise CAIA systems. As Dr Ken Bloom explains, “These systems are meant to complement pathologists, not to replace them.”13 Adopting a virtual workflow-centric system (ie, whole slide scanner) that does not overwhelm the local network would eliminate the need to standardize different acquisition systems and could potentially show better results. New CAIA systems need to offer automatic region-of-interest selection of pertinent areas for rapid analysis. The systems also need to improve based on feedback from pathologists (ie, learning algorithms). Ideally, CAIA needs to be integrated with the laboratory information system to improve workflow. Finally, although CAIA represents a major step in standardizing the quantification of IHC slides, clinical outcome studies are needed prospectively to evaluate this emerging technology. To date, we are aware of at least one study in which investigators showed that CAIA for ER IHC yielded results no different than manual scoring in terms of patient outcome.14Radiology and pathology are moving closer together because the technologic advancements in one field are often applicable to the other. This does not necessarily mean that today's pathologist will be tomorrow's radiologist, or vice versa. It simply means that there is a convergence on the type of tools used in both fields. Pathologists are taking advantage of intelligently designed picture archive and communication systems, which have been used by radiologists to revolutionize workflow. As pathology becomes more digitized, pattern recognition algorithms promise to improve accuracy, reliability, specificity, and productivity. The end result will be an increased reliance on diagnostic tools for earlier identification of specific treatable conditions, and much of this reliance will be focused on where the definitive diagnosis is accomplished. When pathology departments are fully digitized, pathologists will instinctively use computer-assisted, image-based, analytic tools to assess and diagnose specimens.We thank Christopher N. Otis, MD; Giovanna M. Crisi, MD, PhD; Andrew Ellithorpe, MHS; Peter Marquis, BA; and BioImagene, Inc, Cupertino, California.