Feature-based multi-resolution registration of immunostained serial sections
Graphical abstract
Introduction
On the mesoscopic scale, structures like blood vessels or nerves may span over millimeters in biological specimens and thus pose a significant problem for medical imaging. In this context large dimensions are as important as fine details. Such details can only be revealed by microscopy. We propose a “sparse” registration process to solve a fundamental problem concerning 3D reconstruction in microscopic anatomy on a mesoscopic scale. This problem is relevant in immunohistological investigations necessitating serial sections.
Immunohistological staining (immunostaining) is the technique of choice to discriminate cell types in human organ specimens. However, the staining does not penetrate specimens thicker than 50 µm. Such a thickness is also actually too large for transmitted light microscopy. With respect to human specimens, two choices are possible at this point: confocal microscopy or serial sections. Confocal microscopy has the limitations that it cannot be applied to thicker sections and that the sections cannot be stored for longer periods of time. The thickness of a single serial section is 5–10 µm. These sections can be easily stored, can be repeatedly inspected, and allow good penetration of the immunohistological staining solutions. The resolution of a common scanning microscope is 0.25–0.3 µm/pixel. The only drawback is that the limited information in a single section prevents a mesoscopic overview. However, with the modern advances in conventional scanning microscopes, series of high-resolution digital images consisting of up to several hundreds of single sections can be acquired. These images need to be registered. The special problem of this registration is the non-linear distortion of the sections. These distortions are inevitably associated with the sectioning process, because the embedding materials for the organ specimens are relatively soft. This happens in every immunohistological serial section and necessitates the matching of large and small landmarks. We use “sparse,” off-the-shelf feature detection combined with novel matching and warping in non-rigid transform instead of image-based methods and minimize the global distortion over the whole series of sections.
This work focuses on multi-resolution non-rigid registration of serial sections. “Multi-resolution” applies to the varying size of the features used, not to an image pyramid. The non-rigid registration is only a part of the larger production pipeline consisting of image acquisition, registration, segmentation, 3D reconstruction, and visualization.
We present an approach for a “sparse,” feature detection-based multi-resolution registration of microscopic images. Feature-based rigid registration does not pose a significant problem, but it is not sufficient for our purposes, as already detailed earlier (Ulrich et al., 2014). The same publication introduced a section-wide spline-based non-rigid registration that allowed to successfully register complete serial sections at their maximal available resolution.
However, further processing stages uncovered a central problem of “simple” non-rigid registration. Previously, we used only large features. Now, we iterate over multiple feature sizes in the non-rigid registration part. We register the ROIs after a coarse non-rigid registration using the method Ulrich et al. (2014) used for human spleen. We stress that using the “first stage” method alone, or even iterating it again on a ROI yields worse results than our new method. Section 5 details on this.
This paper presents the iterated method consisting of feature detection, matching, filtering, and “warping.” The latter is the process of adapting the B-spline control point grid for actual image (un-)distortion based on similarity of features with close spatial position. This is done at once for all images in the stack. The final step, applying the computed transformation to the actual images, is quite straightforward and is not regarded here in detail. We show the final result of our imaging pipeline in Fig. 9. We compare our approach to the state of the art (Ulrich, Lobachev, Steiniger, Guthe, 2014, Klein, Staring, Murphy, Viergever, Pluim, 2010) using visual methods like volume renderings and z-stacks (Fig. 7, Fig. 8, Fig. 9, Fig. 10, Fig. 11). Further, we quantitatively evaluate our method and related approaches using seven quality measures (Fig. 12, Fig. 13, Fig. 14, Fig. 15).
Investigating micro-vessel density and distribution is relevant in a wide range of experimental therapies, especially in tumor treatment. Fully-automatic image registration on mesoscopic and microscopic scale is instrumental in such research. The main contribution of our paper is the registration based on a pyramid of decreasing sizes of features with a gradually refined control point grid. The “search radius” for spacial location of the features also declines. The result is the non-rigid registration process that fits not only the large features, but also smaller ones. Our larger features are still too small to be the large landmarks, required for, e.g. the method of Bagci et al. (2012). Smaller features represent smaller blood vessels and capillaries, or other microscopic-scale entities of the serial sections. Our registration and undistortion of serial sections allow to answer current research questions like:
- •
“What is the branching pattern of larger blood vessels?”
- •
“Where is the capillary network localized and do regions without capillaries exist?”
- •
“Do special cells exist in the vicinity of the capillaries?”
Section snippets
Related work
Early work on image registration and its applications dates back more than 30 years (Lucas and Kanade, 1981). Friston et al. (1995) describe the image registration process as a least-squares problem where the differences between images are minimized. A survey on early methods was made by Brown (1992). A more recent one by Zitová and Flusser (2003) includes more advanced image registration methods.
Shams et al. (2010) survey methods for medical registration based on local differences.
Outline of the registration process (Fig. 1)
The digitized serial sections are pre-processed to enable better registration. We normalize the sections (Reinhard, Adhikhmin, Gooch, Shirley, 2001, Khan, Rajpoot, Treanor, Magee, 2014). (Depending on the specimen we may also use only a particular image channel for feature detection.) The actual registration procedure can be divided into three phases: a) feature detection, b) rigid registration, c) non-rigid registration. All steps except the second one are repeated allowing for an increasingly
Results
First we state the parameters of the registration we used to obtain the results below. Then we present the specimens that served as inputs to our method. In Section 4.3 we present volume renderings of those specimens.
We show the whole specimens (Fig. 4, Fig. 5, Fig. 6) and registration results of regions of interest as volume renderings and as horizontal z-profiles, i.e. as a side views of the overlaid stack of images. Proper registration is absolutely mandatory to trace the vascular network in
Discussion
We utilized feature detection and matching to non-rigidly register serial sections in a multi-resolution manner. Precise registration allows for a 3D reconstruction of the micro-vasculature. Our results show a substantial increase of visual and quantitative quality over a non-multi-resolution feature detection-based method (Ulrich et al., 2014). Further, our results are superior to the output of the standard multi-resolution image-based method (Klein et al., 2010), and our method is faster. We
Conclusion
This work proposes a new multi-resolution approach to the registration of 2D images. It is used for the analysis of the vascular network obtained from serial sections stained with immunohistological methods that require thin sections. We were able to observe blood vessel networks of various organs in larger areas at a microscopic resolution. The sectioning procedure induces a distortion on the serial sections, which needs to be abolished through our non-rigid registration.
We demonstrate our
Acknowledgments
The authors would like to thank Anja Seiler and Kathrin Lampp from Institute of Anatomy and Cell Biology of Philipps-University Marburg for producing the sections and performing the immunostaining. In addition, we thank Sandra Iden, Nikolay Kladt, and Astrid Schauss from the CECAD Cluster of Excellence at University Cologne for helping a lot with the image acquisition. We thank Norishige Fukushima for his implementation of SSIM. We used MI implementation written by Jose Delpiano. We thank
Dr. Oleg Lobachev holds a MSc in Mathematics (2007) and a PhD in computer science (2011). Oleg was a post-doctoral researcher at Scuola Superiore Sant’Anna in Pisa, Italy; since fall 2012 he is post-doctoral researcher at University Bayreuth. His research interests include 3D reconstruction, medical imaging, GPGPU computing as well as high-level parallel programming and performance estimations.
References (66)
- et al.
A method for registration of 3-D shapes
IEEE Trans. Pattern Anal. Mach. Intell.
(1992) - et al.
Hybrid feature-based diffeomorphic registration for tumor tracking in 2-D liver ultrasound images
IEEE Trans. Med. Imaging
(2013) - et al.
3D volume reconstruction of a mouse brain from histological sections using warp filtering
J. Neurosci. Methods
(2006) - et al.
A nonlinear mapping approach to stain normalization in digital histopathology images using image-specific color deconvolution
IEEE Trans. Biomed. Eng.
(2014) - et al.
A hierarchical approach to elastic registration based on mutual information
Image Vis. Comput.
(2001) - et al.
A multi-resolution approach to non-rigid registration of microscopy images
(2012) - et al.
Reconstructing a 3D structure from serial histological sections
Image Vis. Comput.
(2001) - et al.
WBIR ’14
(2014) - et al.
Numerical Recipes: The Art of Scientific Computing
(2007) - et al.
SIFT and shape context for feature-based nonlinear registration of thoracic CT images
FREAK: Fast Retina Keypoint
KAZE features
Automatic best reference slice selection for smooth volume reconstruction of a mouse brain from histological images
IEEE Trans. Med. Imag.
Hierarchical scale-based multiobject recognition of 3-D anatomical structures
IEEE Trans. Med. Imaging
Registration of standardized histological images in feature space
Proc. SPIE
Speeded-up robust features (SURF)
Comput. Vis. Image Underst.
A survey of image registration techniques
ACM Comput. Surv.
A feature-based, robust, hierarchical algorithm for registering pairs of images of the curved human retina
IEEE Trans. Pattern Anal. Mach. Intell.
Elastic registration of multimodal prostate MRI and histology via multiattribute combined mutual information
Med. Phys.
n-SIFT: n-dimensional scale invariant feature transform
IEEE Trans. Image Process.
Smoothness-guided 3-D reconstruction of 2-D histological images
NeuroImage
Non-rigid image registration: theory and practice
Br. J. Radiol.
Biological imaging software tools
Nat. Methods
Two-frame motion estimation based on polynomial expansion
Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography
Commun. ACM
LSMR: An iterative algorithm for sparse least-squares problems
SIAM J. Sci. Comput.
Spatial registration and normalization of images
Hum. Brain Mapp.
Three-dimensional reconstruction of tumor microvasculature: simultaneous visualization of multiple components in paraffin-embedded tissue
Angiogenes.
Three-dimensional (3D) reconstruction and quantitative analysis of the microvasculature in medulloblastoma and ependymoma subtypes
Angiogenes.
Robust point correspondence applied to two- and three-dimensional image registration
IEEE Trans. Pattern Anal. Mach. Intell.
Feature-constrained nonlinear registration of lung CT images
Med. Image Anal. Clinic
Comparative evaluation of binary features
A global optimisation method for robust affine registration of brain images
Med. Image Anal.
Cited by (25)
A Survey of Methods for 3D Histology Reconstruction
2018, Medical Image AnalysisCitation Excerpt :This process was repeated if the pairs after transformation were not stable. The methodology was reused in Lobachev et al. (2017) and incorporated into a multi-resolution framework. (i) The Euclidean distance in combination with other criteria, like the size of structures, is used to assess the similarity between pairs of such features.
The human splenic microcirculation is entirely open as shown by 3D models in virtual reality
2022, Scientific ReportsInspection of Histological 3D Reconstructions in Virtual Reality
2021, Frontiers in Virtual Reality
Dr. Oleg Lobachev holds a MSc in Mathematics (2007) and a PhD in computer science (2011). Oleg was a post-doctoral researcher at Scuola Superiore Sant’Anna in Pisa, Italy; since fall 2012 he is post-doctoral researcher at University Bayreuth. His research interests include 3D reconstruction, medical imaging, GPGPU computing as well as high-level parallel programming and performance estimations.
Christine Ulrich has a MSc in computer science from Philipps-University Marburg. Her research interests include marching cubes algorithms, 3D-reconstruction, medical imaging, GPGPU computing, and virtual reality. Since fall 2013 she is setting up a VR laboratory at the Institute of Psychology of Philipps-University Marburg.
Prof. Dr. Birte S. Steiniger graduated from Hanover Medical School in 1979 and earned her MD in 1982. She spent her postdoc time in the Department of Anatomy of Hanover Medical School conducting research in lymphatic organs and experimental organ grafts. Having received the venia legendi in 1988, she was appointed to a professorship in the Institute of Anatomy and Cell Biology of Marburg University in 1993. Her research is focused on the microanatomy of human lymphatic organs and on cells of the monocyte-macrophage system.
Dr. Verena Wilhelmi holds a diploma in biology (2006). She received her PhD with studies in particle research at the Heinrich-Heine-University Düsseldorf (2013). Since 2013 she is a post-doctoral researcher at the university of Marburg, Institute of Anatomy and Cell Biology. She is interested in exploring the micro-anatomy and function of structures in spleen an bone marrow by immunohistological methods. By developing a special method to get serial bone slices, this technique allows for 3D reconstruction of blood vessel architecture.
Prof. Dr. Vitus Stachniss graduated from Bonn Dental School in 1974 in dentistry and received his DDS in 1974. He spent his post doc time in the Department of Operative Dentistry in Bonn until 1976, then in Göttingen Dental School until 1983, where he received 1981 the venia legendi (PhD) for Dentistry. He was appointed as chairmen for Operative Dentistry at Marburg University Dental School in 1983 and emeritated in 2008. Since 2005 his research is focused on hard section technology of undecalcified teeth and bone combined with hyperbaric acrylic 3D-orientated embedding.
Prof. Dr. Michael Guthe has obtained his PhD from Rheinische Friedrich-Wilhelm’s-University Bonn in 2005. He obtained his venia legendi in 2012 in Marburg. Since 2012 Michael is Professor at University Bayreuth. He researches geometric modeling, visualization, and parallel programming with focus on GPGPU computing.