 |
|
|
|
|
|
 |
 |
 |
 |
 |
 |
 |
||
Volume 31 |
February 2009 |
Number 1 |
|
||
|
||
Quantitative Histopathology
The Evolution of a Scientific Field Our journal, AQCH, is entering its 30th year of publication. Founded by the late Dr. George L. Wied in 1979, it has become one of the premier journals in the field of quantitative cyto- and histopathology. The 30th anniversary seems like a good occasion to express the editors’ gratitude to our readers, to our contributors, to our tireless reviewers, to the editorial board members and last, but not least, to our publisher, Ms. Donna Kessel of Science Printers and Publishers, Inc. It is also a good time for some reminiscence, to reflect on our beginnings, what we have learned, what worked and what did not, and where we can expect to go from here.
The evolution of a new scientific field is a historic event. There is an element of inevitability, as many precedents confirm. Diligent subjective observation is first supplemented by measures and numeric data, and, inevitably, objective procedures evolve as the basis for the new field. It happened millenia ago in astronomy, in architecture, in maritime navigation, in physics, in chemistry and more recently in sociology and medicine. We have been privileged to be able to witness the early development of quantitative diagnostic histopathology. Not only that, our contributors have been active participants in creating this new scientific field. Creating the field is quite literally a correct term. There was no road map for how to make the diagnostic image information subject to exact measurement. The process has been far from straightforward. Diagnostic histopathology has, for well over a century, been practiced—with eminent success—as an art, firmly founded on the personal experience and judgment of professionals with the highest standards. It is small wonder then that the early beginnings of computer analysis of cytologic and histopathologic imagery were met with a great deal of skepticism. Those of us who have been in this field for almost half a century well recall the then-prevailing opinion that the exquisite capabilities of the eye-brain complex could never be replaced by a “mere machine.” Not that this has ever been the objective of quantification. Nevertheless, to emulate the perceptive capabilities of experienced professionals by analytic procedures is a formidable challenge indeed. There really was no obvious set of procedures to apply. Such measures as nuclear size, nuclear/cytoplasmic ratio and total DNA content clearly were of value but also not sufficient for an unequivocal characterization of individual cells.
Analytic approaches had to be borrowed from other disciplines and adapted. There was literature in pattern recognition. For scene segmentation, edge detection and thresholding in imagery had been found useful in the analysis of reconnaissance imagery. The gray value distribution of digitized images suggested suitable thresholds for the segmentation of cytoplasm and nuclei. Statistical
parameters, such as the mean pixel gray value, standard deviation and curtosis, were used as “features” in cytology.
The next step had its origin in methods used in linguistics to establish authenticity of manuscripts: just as the frequency distribution of different letters and letter sequences had been employed in the identification of authorship, so were pixel gray values and sequences along the scan line used to quantify the nuclear chromatin pattern. The resulting features—relative frequencies of occurrence, co-occurrences, run lengths—have been eminently successful. They have, for decades now, been used to impose a mensuration on nuclear chromatin and are in almost universal use. These features allow a sensitive detection of minute changes in the state of a cell’s differentiation, reflecting, as we are beginning to learn, epigenetic events affecting the gene activation pattern.
They also, though, clearly proved that the computer assessment of digitized imagery allows the extraction of entirely novel, diagnostically useful information, which is visually not reliably, or even not at all, perceived. Computation thus provides a genuine expansion of our ability to perceive. This has led to the discovery of preneoplastic lesions in histopathologically normal-appearing tissue of organs harboring a premalignant or malignant lesion. Preneoplastic lesions have also been documented in patients free from any potentially progressive lesion but at high risk for the development of invasive disease.
The “subvisual” detection capability has proven invaluable as well in the assessment of chemopreventive intervention effects. Such intervention is most appropriate when employed at the earliest stages of deviation from normal. Then, though, visual histopathologic evaluation may be unable to establish a significant effect, whereas karyometry allows statistical documentation.
None of these procedures could have been implemented without multivariate statistical methodology, which has become an indispensable basis for quantitative histopathology.
From pattern recognition research came applications of classification strategies and algorithms. From numerical taxonomy discriminant analysis was obtained, as well as clustering techniques, used in unsupervised learning.
The late 1960s saw much emphasis—and high hopes—for the development of “artificial intelligence.” In our field it ultimately resulted in diagnostic expert systems, in the exploration of methods for uncertainty management in diagnostic decision-making, and in the implementation of artificial neural networks.
The great appeal of these approaches lay in the potential for the integration of linguistic, traditional descriptive diagnostic terms and quantitative numeric variables in the diagnostic process.
Diagnostic expert systems offered some assurance that all potentially useful diagnostic clues would be considered. There were several promising developments. However, in retrospect one finds that it may be unrealistic to expect a histopathologist in clinical practice to answer a long series of questions in order to obtain a diagnostic suggestion. Diagnostic expert systems just do not seem to have caught on in practice. The cause most likely was the human-machine interface. A dialogue mode might have been more appealing. However, the exhaustive collection of all possible considerations to ensure a reliable process proved valuable as the core of knowledge-based, fully automated scene analysis in machine vision. This is where an expert system, as guidance controller, may play a truly valuable role.
The integration of traditional diagnostic linguistic terms and numeric measures was also achieved in belief networks, with some success. A problem presented itself not in the correspondence of a fuzzy linguistic term and a numeric clue value, but in the users’ interpretation of a linguistic term. What was needed was the inclusion of relevant image information as well, i.e., the presentation of sample imagery—possibly with computer graphic enhancement—outlining the range of expression of a linguistic criterion. Today, the availability of huge image databases and extremely high data transfer rates makes this triple integration of linguistics, numeric variable values and image examples feasible.
The San Diego Conference on Neural Network Technology in 1987 raised many hopes for fundamental advances in quantitative histopathology. Neural networks, intuitively, would appear to be a natural methodology. There is no need to define features, no need to develop a computer program. The network would “learn” just like a student of histopathology, by looking at a large number of images of known diagnostic category. Recognition of a “pattern” would develop inevitably. Also, there is another reason why a large neural network would be a natural for image analysis in histopathology. In human vision an entire scene—or microscopic field—is taken in simultaneously. The histopathologic pattern, formed by the histologic components, is immediately apparent. This would be the same for a neural net, much in contrast to conventional computer analysis, where pixels are taken in sequentially and a “pattern” has to be inferred and constructed from such a scan. If one considers all of this, the use of a neural net system in quantitative image analysis has tremendous appeal indeed. Unfortunately, it was soon realized that the idea of a neural net for the analysis of high-resolution imagery was entirely impractical—at least for the time being. The input layer would have to be huge—equal to the number of pixels in the scene—and the hidden layers would require multiples thereof in the number of processing elements. The image of just a single nucleus is sampled typically at a rate of 6 pixels per linear micron. For most nuclei this results in several thousand pixels. A typical network will require about 4 processing elements per input element; i.e., even a single hidden layer network might require well over 20,000 processing elements—just for the recognition of a single nucleus.
Even so, histopathologic imagery poses an additional and more critical problem. The histopathologic pattern is not a topographic one but is expressed as topology. It is expressed in characteristics such as “surrounded by,” “adjacent to,” “extending along,” i.e., in terms of relative positionings. Such a pattern representation is invariant to location, orientation, distortion. In contrast, a topographic pattern is subject to substantial variations. The same problem has been resolved in quantitative cytology and karyometry. Feature extraction there leads to a representation that is invariant to location and orientation.
A neural net would have to map a scene first onto a topologic representation, but such schemes have not yet been explored.
Nevertheless, artificial neural nets have been shown to be eminently useful even when applied only to a set of feature values and have been used to classify nuclei on the basis of a set of karyometric features. They are, of course, in essence, statistical classifiers, but they have the capability to draw very convoluted decision boundaries. One still will have to limit the number of features. An artificial neural net is still subject to the restrictions imposed by the sample size to dimensionality ratio.
Overall, though, one could not say that neural nets have found widespread application in quantitative histopathology. It would seem, though, that the use of Kohonen-type networks in unsupervised learning might offer real advantages, e.g., in knowledge discovery processing and in problems probing the heterogeneity of nuclear populations in phenotypes.
A very substantial number of research efforts have been devoted to the automation of some diagnostic task. The first major effort in this direction was aimed at the automation of screening for cervical cancer. The Cytoanalyzer—for all the negative comments that have been made concerning this project—was technically a remarkable development, considering the state of technology 50 years ago. Its failure to be clinically useful taught an important lesson. Any development in our interdisciplinary field has to have a firm foundation in a thorough understanding of clinical practice and expectations by all involved.
The continued efforts to automate screening for cervical cancer, though, have led to a historic first in medicine. Here, a diagnostic decision—namely, that a clinical sample is within normal limits (WNL)—is made by a machine, and no human professional is expected consequently to look at this sample.
Any automation of analysis and interpretation of histopathologic imagery necessarily involves machine vision. Machine vision systems today are in widespread use in industry, in automated inspection stations, in material handling, in automated assembly. However, one must be aware that these systems function in a very strictly restricted and regulated world. They are dealing primarily with man-made objects of defined morphology and structure, often presented in a prearranged orientation and with known topography.
Histopathologic sections present a much more difficult situation. There is some underlying pattern in tissue of a given organ site. But, as mentioned above, it is characterized by its topologic properties, by the positionings of tissue components relative to each other. These characteristics remain invariant to the manifold morphologic appearances.
It is important to consider what functionality one expects from a machine vision system. A correct segmentation of a scene into its histologic components, their identification, their mensuration, even the assignment of a section to a diagnostic category—all of this can be accomplished by a machine vision system guided by a knowledge base. A knowledge base for control of such a procedure requires “image understanding.” Here, we have an example of canned human intelligence, encoded for use by a computer system to perform an intelligent function. It is not artificial intelligence, though.
It is well understood that a machine vision system designed to face tasks less rigidly defined than those in industrial inspection require, in addition to the imagery, information from other sources. The process is known as “sensor fusion.” It is, in the machine vision environment, nothing else than what has been used in diagnostic histopathology from its beginning, namely, folding in additional knowledge from nonimage sources such as anamnestic data, knowledge of pathologic processes and experience from clinical outcome of similar cases.
It follows, therefore, that “image understanding” by itself would not be sufficient for a diagnostic system. The image understanding capability must be complemented, and in fact, be evaluated and controlled, by systems that one might call “pathology understanding systems,” which provide a knowledge base and fusion of information from various sources.
Development in such a direction has begun with case-based reasoning (CBR) systems. Here, a multitude of data covering all aspects of potential effect on the outcome of a given case are offered in a very large database. The prognostication for a given case is then based only on the known outcomes from a set of the most similar cases. Case-based reasoning is both in principle, and practically, an ideal solution. CBR might be able to cope with the ever-present uncertainty of case-to-case heterogeneity. The problem is that it will require standardization of a wide variety of databases and on an extremely large number of cases to be fully effective.
The evolution of diagnostic histopathology into a quantitative, measuring science is only at its very beginning. It is clear that it takes an integrated effort by an interdisciplinary research community. The search for methodologies that are practical and applicable to the unique problems posed by this field is still on. But, in retrospect and considering from where we started, we have come a long way.
Peter H. Bartels, Ph.D., F.I.A.C.(hon)
From the Arizona Cancer Center, University of
Arizona, 1515 North Campbell Avenue, P.O. Box 245024, Tucson, Arizona 85748, U.S.A., (hubertbartels@msn.com), and
Rodolfo Montironi, M.D., F.R.C.Path.
Section of Pathological Anatomy, Polytechnic University of the Marche Region, Torrette, Ancona, Italy.
Financial Disclosure: The authors have no connection to any companies or products mentioned in this
editorial. Keywords: analytic cytology, quantitative histology. (Anal Quant Cytol Histol 2009;31:1–4)
 0884-6812/09/3101-0001/$18.00/0 © Science Printers and Publishers, Inc.
Analytical and Quantitative Cytology and Histology® Send website comments to Webmaster@aqch.com |
||