Breast cancer is the most common cause of cancer in women worldwide. Cancers of the breast kill more women than any other form of cancer in all parts of the developing world. An important tool for the detection and management of breast cancer is analysis of tissue samples under the microscope by a pathologist. Looking at the cancer cells under the microscope, the pathologist looks for certain features that can help predict how likely the cancer is to grow and spread. These features include the spatial arrangement of the cells, morphometric characteristics of the nuclei, whether they form tubules, and how many of the cancer cells are in the process of dividing (mitotic count). These features taken together determine the extent or spread of cancer at the time of diagnosis.
|Figure 1. Sample images taken from whole slide images of breast tissue. (a) Hematoxylin and eosin (H&E) stained breast specimen containing invasive ductal carcinoma. (b) Normal breast tissue structure|
Visual microscopic interpretation of tissue sections is laborious and prone to subjectivity. Computer-aided diagnosis (CAD) has a huge potential in alleviating shortcomings of human interpretation and will reduce the workload of the pathologists. As a result, more accurate diagnostic information may be extracted, helping clinicians in selecting the most optimal treatment for individual patients. CAD can facilitate diagnosis by sieving out obviously benign slides and providing quantitative characterization of suspicious areas. The aim of this project is to develop image analysis techniques to detect and characterize breast cancer in whole slide images (WSI). Ideally, we will try to extract clinically relevant information related to disease prognosis for individual patients. In particular, methods will be developed to: (1) Standardize the staining of whole slide histopathological H&E slides to facilitate computer aided diagnosis; (2) Automatically detect and segment potential regions of interest in breast tissue; (3) Classify the identified regions as normal or cancerous; (4) Extract meaningful quantitative features, describing cell and tissue characteristics (including visually imperceptible features); (5) Classify the malignant areas into different cancer grades.
Variations in staining color and intensity complicate quantitative tissue analysis. Such variations are due to inter-patient variation and inconsistencies in the preparation of histology slides (e.g. staining duration, stain concentration, tissue thickness). However, staining variations, which may potentially hamper usefulness of computer assisted analysis of histopathological images, can be reduced considerably by applying our proposed algorithm . Figure 2 shows the result of the standardization by the our method.
|Figure 2. Standardization of H&E stained histopathology images. The images on the first row are standardized to have the same spectral characteristic as the image in the bottom right corner.|
Automated detection of clinically meaningful Regions of Interest (RoIs) in WSIs is an essential initial step in the development of an automated computer-aided diagnosis system. Accurate extraction of these RoIs would limit application of complex image analysis tasks only on specific relevant areas within the WSI. Our proposed algorithm  generates and classifies superpixels at multiple resolutions in whole slide histopathology images to detect regions of interest.
|Figure 3. Illustration of multi-scale superpixel classification. (a) Generation of superpixels in multiple levels. (b) Likelihood map showing the probability of a superpixel to belong to the epithelium class. Larger superpixels were classified with high confidence in low magnification. Smaller superpixels are generated and classified at the highest magnification.|
Our current research focus is on automated recognition and grading of tumor. Methods will be developed to classify the identified regions of interest into tumor and normal regions. Subsequently, we will develop algorithms to estimate tumor aggressiveness and predict potential patient outcomes.
This project is part of the European Union FP7 funded VPH-PRISM project and is executed in close collaboration with Fraunhofer MEVIS, Bremen, Germany.
The current project has received financial support by the European Union FP7 funded VPH-PRISM project under grant agreement n◦601040.