The objective of this study was to design and develop an online software application that presents tissue microarray images and stores associated review and clinical data. The result was the Virtual Tissue Matrix (VTM), which consists of TMA images available at multiple magnifications, scoring forms to gather TMA review data and a relational database to store the generated results.
The VTM displays virtual TMA images via a web site and facilitates the storage of TMA data via a relational database. There are numerous advantages of using the VTM over other proposed software systems of its type. Downloading of the images is rapid. Only views that are requested by the user are returned at maximum resolution, thereby downloading the minimum required dataset. The VTM was designed in consultation with scientists and pathologists and, as a result, the reviewing process emulates the workflow involved in conventional TMA reviews. The VTM interface is delivered in HTML, via a conventional web browser, allowing for intuitive user interaction. The VTM database is relational; a structure more suited to the storage of the object oriented dataset generated from TMA experimentation, than previous efforts incorporating flat files and spreadsheets for data storage.
Since the creation of the VTM there have been numerous advances in the technologies used for image acquisition [9
] and image analysis techniques and applications have been well documented in literature [12
]. Integrated intuitive systems are now available that rely on minimal human intervention when scanning slides such as Aperio or Dmetrix [18
]. Numerous commercial image acquisition applications are now available[19
]; however, cost of purchase is often high for these integrated systems putting them out of reach for many research laboratories.
The VTM has been upgraded to support images generated by an Aperio Scanscope T3 Scanner™. Advantages of using the Aperio Scanscope T3 Scanner™ include, batch uploading of slides, the ability to scan glass TMA slides at 20 × magnification within minutes, one touch scanning which reduces manual intervention, automatic section of autofocus points within the tissue, and seamless images with no tiling artifacts. Despite the advances in image acquisition and the obvious advantages automated systems have over older more labour intensive systems, these systems do not wholly address the problem of relating TMA images to review and image analysis data.
The method used to acquire digital images within the VTM is not the main concern of this manuscript, the technology is constantly developing and advancing, and as new and improved systems are developed they can be integrated into the VTM with ease as illustrated by the upgrade to Aperio™.
Once developed the VTM was validated, via assessment of inter-and intra-observer variability on two users' evaluations of immunohistochemically stained tissue microarrays, using digital and microscope analysis. Eight parameters were evaluated, the amount of core and tumour present, the amount and intensity of membrane, cytoplasmic and nuclear staining.
Comparisons evaluated in this study illustrated that intra-and inter-observer virtual TMA reviews produced equivalent levels of agreement as intra-and inter-observer glass TMA reviews, for five out of the eight parameters examined. Where discrepancies occurred it was dependent on the parameters and users involved. In all comparisons, low levels of agreement for the amount of tumour present were observed. This was not surprising, as the application of classifiers to any data continuum (data that does not naturally fall into discrete clusters) results in scoring variability around the interfaces of the classifier. This variability is increased when the number of classes are increased creating more interfaces. Also, of the two reviewers used, one was a scientist and one a pathologist. The scientist accurate interpretation of tumour/non tumour may potentially be questioned as a result of this work.
Of particular interest, were a large number of observations that were considered positively stained by virtual TMA reviews which were considered negatively stained when reviewed using a microscope. This was particularly evident when quantifying the amount of cytoplasmic staining; where virtual TMA reviews observed substantially more positively stained spots than glass TMA reviews. The additional positively stained spots were largely considered to stain between 1–30% of the tumour area and/or to be weakly stained. This suggests that virtual TMA reviews may be more successful in allowing the identification of small areas of staining and/or where staining intensity is low.
One proposed reason for the identification of staining when using digital images that was not observed with a microscope was the use of correcting adjustments to the image data during the digitising of TMAs. Bulbs used in microscopes have a characteristic tint; in general this is yellow or straw coloured. However, this tint is removed when digitising slides using a corrective algorithm, potentially unmasking weak staining that would otherwise be attributable to background tint. Also, with microscope based analysis, background light is adjusted to best suit each individual spot. When digitising the slides for this study, a constant background light intensity was used to digitise all slides for this study.
Excluding nuclear staining, where positive staining was infrequent, agreement levels were low when examining staining intensity. When using a conventional microscope, in general, users failed to utilise all grades within the classifier to characterise positive staining intensity; the category of moderate staining was repeatedly used when positive staining was observed, particularly for membrane staining. However, with digital reviews, all grades within the classifier were utilised more extensively, which suggests that the review of digital images gives a user more confidence to discriminate between different intensities and that subtle differences in intensity may be easier to detect when utilising digital slides, than when utilising a microscope. This may be due to the standardisation in lighting while preparing the images.
Human observers, while excellent at object classification, are inherently poor at quantifying intensities and areas to any degree of accuracy. Studies have shown that image analysis produces more reproducible results than pathologists for quantifying the intensity of staining, in relation to β-Catenin expression in TMAs for colon cancer [22
]. Image analysis systems may identify subtle differences in staining intensity, which are not quantifiable by a human reviewer, thus leading to the better correlation of expression data to prognostic indicators.