According to World Health Organization, breast cancer is the most commonly diagnosed form of cancer in women worldwide and the most common cause of cancer among women in most countries. A well-known method for evaluation of the prognosis of breast cancer is the Bloom-Richardson grading system (1) where three features 1) percentage of tubule formation, 2) degree of nuclear pleomorphism, and 3) mitotic cell count are scored. This is done based on a pathologist's visual examination of tissue biopsy specimens under a microscope and consequently the chances of human error in the disease diagnostics cannot be eliminated. Researchers in this field have recently envisaged that image analysis algorithms may improve the accuracy and reproducibility of estimation of nuclei size and shape differences of the breast histopathology tissues. Advancements in digital pathology and the advent of fast digital slide scanners have simplified the digitisation of histopathology slides which can be analysed using image-based algorithms. [caption id="attachment_33494" align="alignright" width="300"]breast-cancer-histopathology Fig 1: Segmentation of nuclei using histopathology slides[/caption] Automated segmentation of nuclei for quantitative image-based analysis of hematoxylin and eosin (H&E) stained breast histopathology slides (see Fig 1)) is a complex task due to the appearance of the cancerous nuclei, which are often characterised by irregular shapes, vague boundaries and patchy chromatin distribution. In addition, they often occur in overlapping clusters in comparison with the regular shapes, homogenous interiors and smooth boundaries of the unaffected nuclei. Additionally, the high-grade cancer tissues exhibit vesicular nuclei which show visible nucleoli and non-uniform intensity. The segmentation algorithms become intolerant to these cancerous nuclei and give poor results, especially when they are in extensively overlapping clusters (2). The existing image-analysis methods depend on manual training and cannot be generalised because of their sensitivity to initialisations and their limited inability to segment extensive overlapping. All these problems make nuclei detection in high-grade cancer images an open and challenging problem for medical image analysts.

Image segmentation and methodology

An image segmentation-based methodology was proposed that included tensor voting followed by Loopy Belief Propagation (LBP) on a Markov Random Field (MRF) for nuclei delineation (3). The segmentation framework can be divided into four main steps:
  1. Pre-processing: the input RGB image is reduced to a 2D intensity image using principle component analysis (PCA) for concisely representing colour information. The gradient magnitude and directions are computed for this grayscale image
  2. Nuclei saliency map construction using tensor voting: a tensor-voting framework is used to perform voting along the image gradient directions (perpendicular and parallel), which is useful is useful for highlighting salient features, such as the nuclei boundaries and centres, while suppressing background pixels. A saliency map is produced, which combines voting sticks in all directions and this combined map was extremely useful in identifying even vague cells as the boundaries and centres were accentuated;
  3. Nuclei boundary extraction by loopy belief propagation on a Markov random field: at each nuclei centre or seed point, a window-based search (with a per-defined size) was performed to detect the nuclei boundaries using Markov random field (MRF) formulation. Further to this, the loopy belief propagation algorithm is used. It is iterative in that at each iteration, messages (calculated radii depicting potential boundary point) are passed around the nodes in MRF. Once LBP iteration completes, the belief (cost) calculated gives the best radii of every spoke from the centre, which describes the boundary pixels, by calculating its belief;
  4. Post-processing to remove spurious detection regions: this step can be considered as an error-correcting step to filter out the erroneously detected regions. This is achieved by computing the mean of the pixel intensity distribution within each segmented region. As the overall intensity of nuclei tends to be darker then the background pixels, regions with a mean value above some predefined threshold are assumed to be false alarms and are excluded.
The proposed segmentation framework was tested and evaluated on de-identified and de-linked images of histopathology specimens obtained from the Department of Pathology, Christian Medical College Hospital (CMC). The proposed method was trained and validated on eight representative images of H&E stained breast cancer histopathology sections. These images were of tumour growth captured from biopsy slides of different patients through a digital camera, Leica DFC280, attached to a compound microscope setup at x40 magnification. The images had dimensions of 1024 × 1280 pixels. Additionally, two Whole Slide Images (WSI) of H&E stained breast biopsy slides diagnosed for invasive ductal carcinoma were used in the study. The dimensions of the images were 80784 x 148672 pixels captured at x40 magnification by a Ventana slide scanner and stored in the BIF format. All the images were annotated by the pathologist who provided a nuclear pleomorphism score based on the Bloom Richardson protocol. Three important measures – namely: precision, recall and dice coefficient – were used to evaluate the performance of the segmentation algorithm. Test results show that the proposed method is suitable for nuclei segmentation in high-grade breast cancer histopathology images containing scenes depicting Grade 3 nuclear pleomorphism (cancerous nuclei with marked variations from normal nuclei). The success of this method can be considered as a first step towards automating breast cancer diagnosis using biopsy images. However, it is important to remember that automation has its own drawbacks and cancer cell nuclei are not universally pleomorphic.

Multi-disciplinary approach to tackling breast cancer

[caption id="attachment_33482" align="alignright" width="300"]breast-cancer-histopathology-1 Fig 2 (left): Simulated barnacles and (right) detection of real barnacles with meshing for possible finite element application[/caption] A particularly interesting aspect of this study is the multi-disciplinary approach to a traditionally uni-disciplinary problem. The algorithm successfully applied in detecting breast cancer from histopathology images was originally used for underwater structural damage detection. The damages in underwater structures can be cracks, surface corrosion, bio-fouling, scour etc. which can be successfully detected using image analysis algorithms (4,5). Obviously the internal changes which may not be captured by cameras cannot be detected using image analysis. Additionally, turbidity and low visibility in underwater conditions (4) meant that the algorithm needed to be developed for extracting critical information from uncertain and poor data sources. This particular algorithm was used to isolate barnacles as shown in Figure 2. [caption id="attachment_33580" align="alignright" width="300"]figure-3 CLICK TO ENLARGE Fig 3: Applications of LBP-MRF algorithm (A) breast-cancer nuclei detection and (B) barnacle feature identification[/caption] The surface roughness (depth of barnacle cover) and shapes of the barnacles were detected and quantified here. The steps of the algorithm applied for breast cancer detection and for barnacle identification are presented in Figure 3. The application was further extended to develop a simulated barnacle cover for possible finite-element applications for global hydrodynamic load assessment in relation to biofouling in offshore structural elements, including wharves, pipelines and renewable energy devices. Acknowledgements: The authors acknowledge the support of Dr Joy John Mammen and Dr Marie Therese Manipadam of Vellore Christian Medical College, India in relation to the histopathology slides. The authors also acknowledge the support of SFI-ISCA (Science Foundation Ireland -International Strategic Cooperation Award) program grant no. 12/ISCA/2493. Authors: Michael O’Byrne, Maqlin Paramanandam, Bidisha Ghosh, Robinson Thamburaj, Vikram Pakrashi References:
  1. Elston CW, Ellis IO. Pathological prognostic factors in breast cancer. I. The value of histological grade in breast cancer: experience from a large study with long-term follow-up. C. W. Elston & I. O. Ellis. Histopathology 1991; 19; 403–410. Histopathology. 2002;41(3a):151-. pmid:1757079
  2. Irshad, H., Veillard, A., Roux, L., Racoceanu, D., 2014. Methods for Nuclei Detection, Segmentation, and Classification in Digital Histopathology: A Review— Current Status and Future Potential. Biomedical Engineering, IEEE Reviews in 7, 97-114.
  3. Paramanandan M, O’Byrne M, Ghosh B, Mammen JJ, Manipadam MT, Thamburaj R and Pakrashi V. (2016). Automated Segmentation of Nuclei in Breast Cancer Histopathology Images. PloS One, 11(9), 0162053 S
  4. O’Byrne M, Ghosh B, Pakrashi V and Schoefs F. (2014). Effects of Turbidity and Lighting on an Image Processing based Crack Detection Technique, Civil Engineering Research Ireland (CERI) Conference, Belfast, UK
  5. Pakrashi V, Schoefs F, Memet J B and O’ Connor A. (2010). ROC Dependent Event Isolation Method for Image Processing Based Assessment of Corroded Harbour Structures. Journal of Structure and Infrastructure Engineering, 6(3), 365-378