Tissue samples
The solution was elaborated from surgical specimens of invasive breast carcinoma, 20 primary tumors and 1 metastatic lymph node, collected since 2011, from patients followed at the François Baclesse Cancer Centre.
Sample preparation
The most representative slice of each surgical specimen is selected during the macroscopic examination. The whole slice is cut into several pieces according to a regular grid of rectangles whose size should fit those of a standard disposable embedding cassette (31 mm × 25 mm × 5 mm). At this stage, each fragment is oriented and carefully labeled according to its localization recorded on a photographic map. After paraffin embedding, each fragment is trimmed into 3 to 5 µm thick section; each thin paraffin section is spread over the glass slide, while carefully avoiding its upturn, and then submitted to standard protocols.
Staining
The histological sections were stained according to the Hematoxylin-Erythrosine-Saffron (HES), Periodic Acid Schiff (PAS) and/or immunostained for hormonal receptors (ER and PR), HER2, proliferation markers (Ki-67 and PHH3), vascular marker (CD31), using automatons and standard protocols of the pathology department.
Image acquisition
Whole slide images of histological sections were digitized at 20× using the ScanScope CS® microscope slide scanner from Aperio Technologies. They were recorded as tiled tiff images.
Image processing for WSI-tiling
Taking into account the large size of the WSI, a two resolution procedure has been developed. A first composite image is built at a low resolution. The user is allowed, at this step, to correct the automatic stitching of the fragments if needed. Then, the final composite virtual slide is assembled at full resolution by reference to the parameters of the previous adjustment done at low resolution.
First, the program extracts a low resolution sub-image (8 µm per pixel) from the WSI pyramid of each fragment; it corresponds to a sixteen fold (x and y) down-sampling of the full resolution image (0.5 µm per pixel). Then it constructs the corresponding untiled binary mask of the tissue and computes its edges. The fully automatic program builds thereafter horizontal strips by bringing together recursively binary images of tissue fragments from the left to the right according to the original map. For each strip, it straightens first the fragment on the left to align its right-side edge pixels; it then straightens the fragment on the right to align its left-side edge pixels. The two fragments are thereafter registered by matching the right-side edge center of the first straightened fragment with the left-side edge center of the second straightened fragment of the tissue section (Figure 1a). Finally, the low resolution composite binary mask is built by assembling recursively downwards the horizontal strips of fragments previously stitched: the application straightens the top strip to align its bottom-part pixels, straightens the bottom strip to align its top-part pixels, then registers both strips by matching bottom-side center of the straightened top strip with the up-side center of the straightened bottom strip (Figure 1b).
A low resolution untiled color image is built in parallel taking the straightened binary masks as a reference, in order to visually evaluate the reconstruction result. At this step, if needed, the user can correct the original binary masks as well as the stitched strips with an image editing software in order to reduce artifacts at the edges. Since the algorithm uses the binary masks including any adjustments, the user can repeat this low resolution correction until a satisfactory reconstruction is obtained.
For the construction of the color high resolution composite virtual slide, each WSI and each strip are straightened and stitched by reference to the distortion parameters of the low resolution binary masks, refined by a polynomial interpolation. The final CVS is provided as a tiled "BigTiff" image.
The program was developed in C language without any parallelization and ran on 32 and 64 bits personal computers under Windows and Linux operating systems.
Quality control protocol
This protocol consists of four steps which were applied on five breast tumors to evaluate the internal deformation of tissue structures. These tumors were chosen in order to get a representative whole section on a single glass slide (HES staining). First, the WSI was split in several pieces (4 to 17). Then, the entire section was reconstructed thanks to WSI-tiling program. Finally, each obtained CVS was compared to the original WSI. Due to the huge size of the images the comparison was done on a representative sample of them. For this purpose, a random systematic sampling (spacing of 2500 pixels) was done on the paired images, WSI and corre-sponding CVS. This sampling allows the extraction of a high resolution gallery of representative thumbnails of 500 × 500 pixels for each image. Each pair of thumbnails was registered [7] and the correlation coefficient was computed between these registered thumbnails. This procedure was reiterated four times.
Another control was done on images with less than 6 pieces to evaluate the influence of dis-tortions by rotation and homothety of each piece.