In this WP, the algorithms developed in WP 1&2 will be combined and optimized for FL image classification and rating. Dr. Gurcan’s group at OSU has already established databases for the development of algorithms, secured institutional review board approvals, created the software infrastructure to classify and rate FL images using supercomputers, which involve the following image processing steps:
- Follicle region detection in IHC images
- registration of IHC images to H&E images
- follicle region detection in H&E images using the registration outcome
- detection of centroblasts and non-centroblasts, and integration and optimization of the overall system.

IHC images have better contrasts between follicle and inter-follicle regions as compared to H&E images; that is why the whole process starts with IHC images. The detected follicle regions in IHC images are then mapped to H&E images where follicle boundaries are further adjusted. Since this is a complex and multi-dimensional problem, it requires close cooperation of image recognition and pattern recognition experts. All the European partners have already attacked similar medical and non-medical problems and have several years of expertise in these areas.

WARWICK has developed a wavelet transform-domain method, which adaptively selects the best wavelet features that are optimal in discriminating between different classes of histology images. The covariance and co-difference matrices constructed from optimally selected wavelet features and MRF segmentation algorithms, will be investigated in both IHC and H&E images for follicular boundary detection.

Similarly, the Quantitative Histo-Imagery team of the GRECAN has been working on segmenting and extracting features from low resolution virtual slides of whole tumor tissue section since 1995. GRECAN recently started using a multi-scale wavelet approach, the team adapts various tools [25] developed for low resolution images in order to process high resolution virtual slides. The first step of the process is subsampling of high-resolution virtual slides acquired at a resolution of 0.25µm, 0.5µm or more (up to the resolution of 4µm). In order to preserve structural details and colours during the subsampling process, an algorithm based on wavelet filtering is used. This allows processing the whole slide at once on a standard PC, in order to detect and localize objects-of-interest at low-resolution. Computation of some features can also be done at this stage, in order to perform a pre-classification of tumors. After that, the boundaries of objects-of-interest can be mapped to the high-resolution virtual slides for visual control. The objects can be copied and pasted into a blank image series to generate a high-resolution gallery. These galleries will then be used for refining segmentation results and extracting high-resolution features.
The joint expertise in multi-scale image analysis and medical image processing will bring computationally efficient solutions to the subject problem and will contribute to Ph.D. theses of young scientists in the MIRACLE group.