Get Complete Project Material File(s) Now! »
Histopathological biomarkers quanti cation
One essential question that arises during histological investigations is how much of a given marker is present in a tissue of interest. As histological processing relies on tissue sectioning, answering this question involves two steps: (1) evaluate the amount of marker in tissue section images and (2) infer the marker quantity at the tissue-level using data obtained from the sampled tissue sections. The rst step has been thoroughly investigated and is still a source of innovations but, interestingly, the second step has been relatively poorly addressed in the literature. Here, we will mostly focus on available methods to address the rst step, while the second aspect will be the topic of section 4.3.3. After reviewing general aspects of histopathological marker quanti cation, we will focus on available methods for pathological markers quanti cation in brain tissue section images.
Contrary to quantitative imaging modalities, such as positron emission tomogra-phy and autoradiography, where the imaging signal is directly related to the quantity of radioactive tracer within the tissue, histology imaging provides contrasts that are not directly related to the underlying amount of marker of in-terest. Depending on the nature of the investigated histopathological marker, 2 di erent approaches can be employed:
Pixel-level quanti cation relates to the fraction of pixels that are considered to be positively stained within the tissue.
Object-level quanti cation consists in detecting objects as groups of con-nected pixels according to some criteria and to evaluate the number of objects within the tissue.
In this work, we will mostly study neuropathological markers of Alzheimer’s dis-ease. Pathological markers are by de nition, not well-de ned objects, thus, the question that we will try to answer relates to how much of the tissue is a ected by pathological markers. Therefore, the pixel-level quanti cation approach is usu-ally more biologically relevant than the object-level quanti cation. On the contrary when studying cellular markers, object-level quanti cation is often more relevant because cells have a clearly de ned structure and are separated from each other by their membrane. Many algorithms have been developed for computer-based cellular object-level quanti cation, mostly in the context of tumor histopathology .
Another important aspect concerns the scale in which is expressed marker quan-ti cation:
Categorical scale: tissue samples are assigned to one of a few ordered cat-egories. For example, the Braak stages consist of 6 categories and are used by pathologists to evaluate the severity of neuro brillary tangle deposition in Alzheimer’s disease . Sometimes this approach is referred as semi-quantitative.
Continuous/discrete scales: in the case of pixel-level quanti cation, the fraction of positively stained pixels is a positive continuous random variable. In the case of object-level quanti cation, object count is a positive discrete random variable.
Preclinical research on Alzheimer’s disease
Substantial knowledge about Alzheimer’s pathophysiology comes from animal stud-ies. A plaque and neuro brillary tangle deposits have been observed in some species including dogs, cats and non-human primates [53, 54] but the presence of Alzheimer’s disease neuropathological markers in animals does not strongly corre-lates with neuronal death or cognitive decline. Nevertheless, in mouse lemurs, a well-studied small non-human primate, brain atrophy and cognitive decline appear in a small fraction of old animals but this is not correlated with A and neuro bril-lary tangle deposition . For ethical and practical reasons, non-human primates are not commonly used as models of Alzheimer’s disease and mouse models have been developed.
As mouse do not naturally develop Alzheimer’s disease, transgenic mouse models have been developed to mimic the pathology. APP, PS1 and PS2 mutations found in familial forms of Alzheimer’s disease have been transferred to mice which resulted in the generation of a number of lines of transgenic mice harboring single or multiple mutations. These mice typically show increased production of A peptide, A plaque deposition and Tau hyperphosphorylation but they do not present neu-ronal death or neuro brillary tangles . Nonetheless, A plaque deposition seems to have a wide range of e ects in these mice including synaptic impairment and moderate cognitive de cits . Also, 3D-autoradiography studies in our labora-tory suggest that transgenic mouse models of Alzheimer’s disease present brain glu-cose metabolism that resemble early Alzheimer’s disease [57, 58]. Triple transgenic mice harboring APP,PS1 and Tau mutations found is rare forms of familial human Tauopathies have been developed in order to model the full range of pathological markers in Alzheimer’s disease. These mice develop A plaques and neuro brillary tangles with time and synaptic dysfunction but only moderate neuronal death and cognitive impairment .
Since 1995, when the rst transgenic mouse models of Alzheimer’s disease be-came available , many therapeutics have shown promising results in preclinical studies but failed in clinical trials. Zahs et al. argued that preclinical studies fail in to translate into clinical applications because transgenic mouse models resemble to early Alzheimer’s disease when patients still do not show signs of cognitive im-pairment . Recently, many of the drugs that have failed to show any e ect in Alzheimer’s disease are being investigated again in patients that show rst signs of mild cognitive impairment. Since A load PET tracers are available, clinicians can detect patients at risk much earlier than the onset of cognitive impairment and pro-pose them to participate in clinical trials with anti-A therapeutics with the hope that these therapies can slow the progression of the disease before brain damage is too important.
Whole-slide imaging strategies
A single histological series typically contains 100 sections per mouse brain. Thus, for group-studies with several histological series, thousands of sections need to be digitized. As mentioned in section 1.1.2, high-throughput histology imaging implies the use of dedicated systems such as atbed scanners and whole-slide imaging mi-croscopes. In the next chapters, both approaches will be used depending on the ap-plication. Here, we will discuss advantages and weaknesses of both systems and how to choose between them. This will be illustrated with results obtained using systems that were available during this work, namely, the ImageScanner III (GE) atbed scanner and the Axio ScanZ.1 (Zeiss) whole-slide imaging mi-croscope.
The rst and obvious di erence between both systems is the resolution. Flatbed scanners resolution is limited by the CCD camera and, in our case, ImageScanner III can image sections at resolutions of 25 m or 5 m. Whole-slide imaging micrscopes use lengths to magnify the image. In our case, Axio ScanZ.1 can image sections with 10 , 20 , 40 magni cations which correspond to resolutions of, respectively, 0.44 m, 0.22 m and, theoretically, 0.11 m. The scope of what can be resolved with both systems is very di erent. The atbed scanner is limited to imaging rather big objects. Brain cells, which vary in size around a few m, cannot be resolved prop-erly by the atbed scanner and studying them will require the use of the whole-slide imaging microscope. Figure 2.4a shows an example where neither individual phago-cytic cells nor individual Nissl-stained cells can be resolved with the atbed scanner. In addition to cellular markers, some intra-cellular neuropathological markers such as neuro brillary tangles may not adequately be resolved as well.
Other types of neuropathological markers such as A plaques or brain blood leakages are clearly discernable with the atbed scanner (Figure 2.4c). However, the amount of detail is much lower than when using the whole-slide imaging microscope and more noise is introduced by the digitization process. We investigated whether these di erences alter A load quanti cation. A set of 96 sections from brains of 12 transgenic APP/PS1 mice stained for A plaque detection (6E10 IHC) were imaged at both microscopic (0.44 m) and mesoscopic (5 m) resolutions. Images were then segmented using BioVision algorithm  and the A load was calculated as the percentage of positively stained pixels within the brain tissue. Figure 2.5 shows that quanti cation at both resolutions is very strongly correlated. While some microscopic details may not be apparent with the atbed scanner, the overall quanti cation is very close to that obtained at high-resolution which indicates that the flatbed scanner is a valid imaging system to study the Aβ load despite the relatively low resolution. Another important aspects are the scanning time and workload. The overall time for scanning 100 sections at lateral resolution of 5 µm is about 2 minutes with our flatbed scanner while, at lateral resolution of 0.44 µm, it takes a few hours with our whole-slide imaging microscope. Flatbed scanner imaging requires less time and workload than when using whole-slide imaging microscopy for the following reasons:
• Lower resolution images are smaller in size and, thus, faster to digitize (it is, however, possible to fairly reduce high-resolution microscopy imaging time by detecting sections on a preview low resolution image and image only these regions at high-resolution).
• Image digitization via line-scan CCD camera is rapid and whole slides images are obtained without tile/line stitching.
• As depth of field reduces with high-resolution microscopy imaging, additional time is needed for fine focus adjustment. Also, sometimes, automatic focusing will fail to find an adequate focal plane and the section will have to be imaged one more time.
Supervised segmentation of neuropathological markers
Neuropathological markers are extensively studied in human brain samples and ani-mal models of Alzheimer’s disease and they represent important therapeutic targets. Thus, their precise quanti cation is a critical issue for both physiopathological re-search and drug development. Quanti cation is commonly performed by segmenting histology whole slide microscopy images and computing the proportion of positively stained pixels relative to the remaining brain tissue.
In this context, global and adaptive thresholding are popular segmentation ap-proaches for they are simple and fully automated (see section 1.1.3). However these methods are prone to errors. Indeed, histological procedures often lead to the pres-ence of artifacts, such as debris and tissue folding, which have similar color properties to the marker of interest. In addition, histological staining can lead to important background nonspeci c staining. This hinders methods based solely on color to provide optimal segmentation results and, as neuropathological markers represent a tiny portion of the brain tissue, minor segmentation errors can strongly impact the overall quantification.
To overcome these limitations, Chubb et al.  proposed BioVision, a super-vised classification approach which uses color and local intensity information (mean of the R, G and B channels in a 16-pixels diagonal neighborhood, section 1.1.3). This latter feature is particularly helpful to account for noise in neuropathology images. For each class, BioVision estimates the joint distribution of the predictors using Gaussian Mixture Models. In the segmentation step, each pixel is classified with a Bayesian classifier. Class imbalance between neuropathological markers and the rest of the tissue can be accounted for by injecting prior probabilities of each class in the decision. Although numerous contextual features have been shown to be eﬃcient for image segmentation, they are typically high-dimensional and because of the curse of dimensionality, density estimation for these features with GMM is hazardous. Thus, one limitation of BioVision is that it cannot adequately incorporate complex contextual features, such as texture descriptors, which could be beneficial to the classification task. The authors proposed to use an automatic post-processing step to remove incorrectly classified pixels based on shape and size features. However, pathological aggregates are not well-defined objects. As shown in Figure 2.10, they present irregular shapes and disparate sizes. Hence, misclassified pixels are hard to detect with morphological features and a robust one-step segmentation approach would be preferable
Table of contents :
1.1.1 Principles and evolution
1.1.2 Histology imaging
1.1.3 Histopathological biomarkers quantication
1.2 Alzheimer’s disease
1.2.1 Impact on public health
1.2.2 Pathological hypotheses
1.2.3 Preclinical research on Alzheimer’s disease
2 High-throughput 3D histopathology
2.1 Large-scale tissue section processing
2.2 Whole-slide imaging strategies
2.3 Brain multi-modal 3D reconstruction
2.4 Supervised segmentation of neuropathological markers
3 Brain-wide quantitative analysis of neuropathological markers
3.1 Ontology-based analysis
3.1.1 3D mouse brain atlases
3.1.2 Atlas registration to histopathological data
3.1.3 Mouse brain ontology
3.2 Voxel-wise analysis
3.2.1 Principle and issues for histopathological data analysis
3.2.2 Heat map generation
3.2.3 Cluster detection with simulated data
4 Applications in mouse models of A deposition
4.1 Context of the studies
4.2 Study 1: animal model characterization
4.2.1 Experimental procedures
4.2.2 Ontology-based characterization of A deposition
4.3 Study 2: preclinical immunotherapy evaluation
4.3.1 Experimental procedures
4.3.2 Whole-brain therapeutic intervention evaluation
4.3.3 3D versus 2D histopathology comparison
4.4 Study 3: gene eect evaluation
4.4.1 Experimental procedures
4.4.2 Image processing
4.4.3 ADAM30 eect on A pathology
5 Towards multimodal and multiscale analysis
5.1 Matching quantitative 3D histopathology to in vivo MRI
5.2 3D whole-brain histopathology at the cellular-scale
5.2.1 Spatial correlation between cellular markers at the wholebrain level
5.2.2 A hierarchical Gaussian Mixture Model (GMM) approach for touching cells segmentation
A 3D Histopathology Analysis Pipeline (3D-HAPi)
B Experimental procedures summary