Key Technologies  

Tissue Microarrays as Biomarker Validation Tools

The development of tissue microarrays (TMAs), wherein hundreds of clinical specimens can be arranged in an ordered fashion on a single tissue section, has allowed for candidate cancer biomarkers to be rapidly and uniformly assessed as to potential clinical utility. Techniques such as IHC or in situ hybridisation (ISH) when applied to TMAs can obtain a simultaneous view of protein or mRNA expression across a wide cohort of patients with different clinical outcomes. For this reason, the approach has been termed “high-throughput”. While this is true, in that it is a single assay platform employed for multiple samples, the subsequent analysis of biomarker expression on TMAs is still based on visual inspection and scoring by a trained pathologist. Whilst critical for successful biomarker analysis using TMAs, it represents a significant bottleneck in many studies. In addition to the time it takes to manually score hundreds of tissue cores, there are also issues associated with inter- and intra-observer reproducibility of scoring due to the subjectivity of visual interpretation by the naked eye. It is for these reasons that computerised image analysis has come to the fore, as a means of supplementing biomarker evaluation by pathologists using TMAs.


Manual interpretation of immunohistochemistry (IHC) is a subjective, time-consuming process, subject to an inherent intra- and inter-observer variability. Automated image analysis approaches offer the possibility of developing more robust and rapid indicators of IHC staining. OncoMark has created a novel, proprietary image analysis toolkit for assessment of IHC-based markers. IHC-MARK, which is in beta-testing phase at present, provides an unparalleled approach to discrimination of biomarker protein expression at the subcellular level (see example output data in adjacent panel).  
Example of automatic assessment of nuclear protein expression in breast cancer tissues. A novel  unsupervised clustering approach was used to overcome one of the major barriers in image analysis -  the identification of tumour nuclei from stromal tissue. The output shows positive tumour nuclei in red and negative tumour nuclei in blue. Data generated from the analysis include % positive nuclei and nuclear intensity. 

Automated Identification of different histological features in prostate cancer tissue

As part of the Irish Prostate Cancer Research Consortium, UCD have developed a unique methodology for computer-aided prostate cancer detection in whole-mount histopathology slides. This automated tile-based technique, which takes advantage of the CIEL*a*b colour space, offers a novel method for high-volume whole-slide image processing and analysis. The UCD group then utilised a random forest feauture selection approach to compare the generalisation nature of CIEL*a*b versus RGB colour texture features to identify the most suitable features for prostate cancer detection. They demonstrated that a classifier performance of 95% is achieved using figures based on random forest, as well as support vector machine approaches.

High Performance Computing

Until recently, image analysis of TMAs would have been impracticable, since recording separate digital images of each individual core using a standard camera would have been enormously time consuming. However, the advent of virtual microscopy and high-resolution scans of entire glass slides has allowed an entire TMA slide to be scanned in a few minutes, completely capturing the biomarker densitometric and location information in the form of a single digital image. This provides an ideal platform to explore the use of computer-based algorithms for automated analysis of tissue biomarkers within TMAs (as well as full-face sections), with the resulting opportunity to develop a truly high-throughput platform for biomarker discovery in tissues. A number of commercial systems are currently available which provide computer-based analysis of TMAs using generic algorithms for nuclear/cytoplasmic segmentation and quantitation of IHC; however, these still tend to be limited by requiring a considerable degree of upfront training/a priori knowledge. As mentioned above, one of the major technical challenges in using virtual slides is the size of the images generated. Scanning a typical region of 25mm×15mm occupied by a standard TMA sample on a glass slide at 40X magnification can result in an image with 100,000×60,000 pixels, corresponding to 20GB of uncompressed data. At this resolution, an individual tissue core of approximately 0.6 mm in diameter would be approximately 9 mega-pixels. Analysing tissue structure and biomarker density in images of this size on multiple cores is computationally intensive and time consuming. However, by analysing multiple cores simultaneously, using high-performance computing (HPC), one could theoretically significantly speed up biomarker quantitation on tissues. The discrete nature of a TMA and its component tissue samples lends itself perfectly to independent and highly parallelised analysis.

Others have considered this concept in the context of Grid-based computing which is a highly distributed form of computing using a decentralised model. Whilst providing certain speed advantages, Grid-based computing can be difficult to control, manage and configure for dedicated experiments. This arises from the fact that it tends to incorporate heterogeneous collections of computers, with widely different capabilities, managed by different organisations, widely distributed geographically, with inconsistent connections and bandwidth. As an alternative approach, dedicated high-performance computer clusters specifically designed for high-throughput analysis of TMAs (and full-face sections) can be implemented. The benefits of cluster-based computing are that the computer architecture can be specifically designed to manage parallel processing with consistency across processors in the cluster and fast connections among nodes. This can provide a convenient and highly rapid approach for automated image analysis of tissues. In conclusion, FAST-PATH is centred on developing and applying automated image analysis focused on key aspects of prostate cancer diagnostics, complemented by a sophisticated HPC framework.