Design-based counting techniques: the real problems

SPECHT: Self-tuning Plausibility based object detection Enables quantification of Conflict in Heterogeneous multi-scale microscopy

Identification of small objects in fluorescence microscopy is a non-trivial task burdened by parameter-sensitive algorithms, for which there is a clear need for an approach that adapts dynamically to changing imaging conditions. Here, we introduce an adaptive object detection method that, given a microscopy image and an image level label, uses kurtosis-based matching of the distribution of the image differential to express operator intent in terms of recall or precision. We show how a theoretical upper bound of the statistical distance in feature space enables application of belief theory to obtain statistical support for each detected object, capturing those aspects of the image that support the label, and to what extent. We validate our method on 2 datasets: distinguishing sub-diffraction limit caveolae and scaffold by stimulated emission depletion (STED) super-resolution microscopy; and detecting amyloid-β deposits in confocal microscopy retinal cross-sections of neuropathologically confirmed Alzheimer's disease donor tissue. Our results are consistent with biological ground truth and with previous subcellular object classification results, and add insight into more nuanced class transition dynamics. We illustrate the novel application of belief theory to object detection in heterogeneous microscopy datasets and the quantification of conflict of evidence in a joint belief function. By applying our method successfully to diffraction-limited confocal imaging of tissue sections and super-resolution microscopy of subcellular structures, we demonstrate multi-scale applicability.

Quantitative neurotoxicology: Potential role of artificial intelligence/deep learning approach

Neurotoxicity studies are important in the preclinical stages of drug development process, because exposure to certain compounds that may enter the brain across a permeable blood brain barrier damages neurons and other supporting cells such as astrocytes. This could, in turn, lead to various neurological disorders such as Parkinson's or Huntington's disease as well as various dementias. Toxicity assessment is often done by pathologists after these exposures by qualitatively or semiquantitatively grading the severity of neurotoxicity in histopathology slides. Quantification of the extent of neurotoxicity supports qualitative histopathological analysis and provides a better understanding of the global extent of brain damage. Stereological techniques such as the utilization of an optical fractionator provide an unbiased quantification of the neuronal damage; however, the process is time-consuming. Advent of whole slide imaging (WSI) introduced digital image analysis which made quantification of neurotoxicity automated, faster and with reduced bias, making statistical comparisons possible. Although automated to a certain level, simple digital image analysis requires manual efforts of experts which is time-consuming and limits analysis of large datasets. Digital image analysis coupled with a deep learning artificial intelligence model provides a good alternative solution to time-consuming stereological and simple digital analysis. Deep learning models could be trained to identify damaged or dead neurons in an automated fashion. This review has focused on and discusses studies demonstrating the role of deep learning in segmentation of brain regions, toxicity detection and quantification of degenerated neurons as well as the estimation of area/volume of degeneration.

On the objectivity, reliability, and validity of deep learning enabled bioimage analyses

Bioimage analysis of fluorescent labels is widely used in the life sciences. Recent advances in deep learning (DL) allow automating time-consuming manual image analysis processes based on annotated training data. However, manual annotation of fluorescent features with a low signal-to-noise ratio is somewhat subjective. Training DL models on subjective annotations may be instable or yield biased models. In turn, these models may be unable to reliably detect biological effects. An analysis pipeline integrating data annotation, ground truth estimation, and model training can mitigate this risk. To evaluate this integrated process, we compared different DL-based analysis approaches. With data from two model organisms (mice, zebrafish) and five laboratories, we show that ground truth estimation from multiple human annotators helps to establish objectivity in fluorescent feature annotations. Furthermore, ensembles of multiple models trained on the estimated ground truth establish reliability and validity. Our research provides guidelines for reproducible DL-based bioimage analyses.

DALMATIAN: An Algorithm for Automatic Cell Detection and Counting in 3D

Current 3D imaging methods, including optical projection tomography, light-sheet microscopy, block-face imaging, and serial two photon tomography enable visualization of large samples of biological tissue. Large volumes of data obtained at high resolution require development of automatic image processing techniques, such as algorithms for automatic cell detection or, more generally, point-like object detection. Current approaches to automated cell detection suffer from difficulties originating from detection of particular cell types, cell populations of different brightness, non-uniformly stained, and overlapping cells. In this study, we present a set of algorithms for robust automatic cell detection in 3D. Our algorithms are suitable for, but not limited to, whole brain regions and individual brain sections. We used watershed procedure to split regional maxima representing overlapping cells. We developed a bootstrap Gaussian fit procedure to evaluate the statistical significance of detected cells. We compared cell detection quality of our algorithm and other software using 42 samples, representing 6 staining and imaging techniques. The results provided by our algorithm matched manual expert quantification with signal-to-noise dependent confidence, including samples with cells of different brightness, non-uniformly stained, and overlapping cells for whole brain regions and individual tissue sections. Our algorithm provided the best cell detection quality among tested free and commercial software.

Three-dimensional and stereological characterization of the human substantia nigra during aging

The human brain undergoes non-uniform changes during aging. The substantia nigra (SN), the source of major dopaminergic pathways in the brain, is particularly vulnerable to changes in the progression of several age-related neurodegenerative diseases. To establish normative data for high-resolution imaging, and to further clinical and anatomical studies we analyzed SNs from 15 subjects aged 50-91 cognitively normal human subjects without signs of parkinsonism. Complete brains or brainstems with substantia nigra were formalin-fixed, celloidin-mounted, serially cut and Nissl-stained. The shapes of all SNs investigated were reconstructed using fast, high-resolution computer-assisted 3D reconstruction software. We found a negative correlation between age and SN volume (p = 0.04, rho = -0.53), with great variability in neuronal numbers and density across participants. The 3D reconstructions revealed SN inter- and intra-individual variability. Furthermore, we observed that human SN is a neuronal reticulum, rather than a group of isolated neuronal islands. Caution is required when using SN volume as a surrogate for SN status in individual subjects. The use of multimodal sequences including those for fiber tracts may enhance the value of imaging as a diagnostic tool to assess SN in vivo. Further studies with a larger sample size are needed for understanding the structure-function interaction of human SN.

How to count cells: the advantages and disadvantages of the isotropic fractionator compared with stereology

The number of cells comprising biological structures represents fundamental information in basic anatomy, development, aging, drug tests, pathology and genetic manipulations. Obtaining unbiased estimates of cell numbers, however, was until recently possible only through stereological techniques, which require specific training, equipment, histological processing and appropriate sampling strategies applied to structures with a homogeneous distribution of cell bodies. An alternative, the isotropic fractionator (IF), became available in 2005 as a fast and inexpensive method that requires little training, no specific software and only a few materials before it can be used to quantify total numbers of neuronal and non-neuronal cells in a whole organ such as the brain or any dissectible regions thereof. This method entails transforming a highly anisotropic tissue into a homogeneous suspension of free-floating nuclei that can then be counted under the microscope or by flow cytometry and identified morphologically and immunocytochemically as neuronal or non-neuronal. We compare the advantages and disadvantages of each method and provide researchers with guidelines for choosing the best method for their particular needs. IF is as accurate as unbiased stereology and faster than stereological techniques, as it requires no elaborate histological processing or sampling paradigms, providing reliable estimates in a few days rather than many weeks. Tissue shrinkage is also not an issue, since the estimates provided are independent of tissue volume. The main disadvantage of IF, however, is that it necessarily destroys the tissue analyzed and thus provides no spatial information on the cellular composition of biological regions of interest.

Neuropathological and Reelin deficiencies in the hippocampal formation of rats exposed to MAM; differences and similarities with schizophrenia

BACKGROUND

Adult rats exposed to methylazoxymethanol (MAM) at embryonic day 17 (E17) consistently display behavioral characteristics similar to that observed in patients with schizophrenia and replicate neuropathological findings from the prefrontal cortex of psychotic individuals. However, a systematic neuropathological analysis of the hippocampal formation and the thalamus in these rats is lacking. It is also unclear if reelin, a protein consistently associated with schizophrenia and potentially involved in the mechanism of action of MAM, participates in the neuropathological effects of this compound. Therefore, a thorough assessment including cytoarchitectural and neuromorphometric measurements of eleven brain regions was conducted. Numbers of reelin positive cells and reelin expression and methylation levels were also studied.

PRINCIPAL FINDINGS

Compared to untreated rats, MAM-exposed animals showed a reduction in the volume of entorhinal cortex, hippocampus and mediodorsal thalamus associated with decreased neuronal soma. The entorhinal cortex also showed laminar disorganization and neuronal clusters. Reelin methylation in the hippocampus was decreased whereas reelin positive neurons and reelin expression were unchanged.

CONCLUSIONS

Our results indicate that E17-MAM exposure reproduces findings from the hippocampal formation and the mediodorsal thalamus of patients with schizophrenia while providing little support for reelin's involvement. Moreover, these results strongly suggest MAM-treated animals have a diminished neuropil, which likely arises from abnormal neurite formation; this supports a recently proposed pathophysiological hypothesis for schizophrenia.

Calibration of the stereological estimation of the number of myelinated axons in the rat sciatic nerve: a multicenter study

Several sources of variability can affect stereological estimates. Here we measured the impact of potential sources of variability on numerical stereological estimates of myelinated axons in the adult rat sciatic nerve. Besides biological variation, parameters tested included two variations of stereological methods (unbiased counting frame versus 2D-disector), two sampling schemes (few large versus frequent small sampling boxes), and workstations with varying degrees of sophistication. All estimates were validated against exhaustive counts of the same nerve cross sections to obtain calibrated true numbers of myelinated axons (gold standard). In addition, we quantified errors in particle identification by comparing light microscopic and electron microscopic images of selected consecutive sections. Biological variation was 15.6%. There was no significant difference between the two stereological approaches or workstations used, but sampling schemes with few large samples yielded larger differences (20.7+/-3.7% SEM) of estimates from true values, while frequent small samples showed significantly smaller differences (12.7+/-1.9% SEM). Particle identification was accurate in 94% of cases (range: 89-98%). The most common identification error was due to profiles of Schwann cell nuclei mimicking profiles of small myelinated nerve fibers. We recommend sampling frequent small rather than few large areas, and conclude that workstations with basic stereological equipment are sufficient to obtain accurate estimates. Electron microscopic verification showed that particle misidentification had a surprisingly variable and large impact of up to 11%, corresponding to 2/3 of the biological variation (15.6%). Thus, errors in particle identification require further attention, and we provide a simple nerve fiber recognition test to assist investigators with self-testing and training.

Optical disector counting in cryosections and vibratome sections underestimates particle numbers: effects of tissue quality

Optical disector counting is currently applied most often to cryosections, followed in frequency by resin-embedded tissues, paraffin, and vibratome sections. The preservation quality of these embedding options differs considerably; yet, the effect of tissue morphology on numerical estimates is unknown. We tested whether different embedding media significantly influence numerical estimates in optical disector counting, using the previously calibrated trochlear motor nucleus of hatchling chickens. Animals were perfusion-fixed with paraformaldehyde (PFA) only or in addition with glutaraldehyde (GA), or by Methacarn immersion fixation. Brains were prepared for paraffin, cryo-, vibratome- or celloidin sectioning. Complete penetration of the thionin stain was verified by z-axis analysis. Neuronal nuclei were counted using an unbiased counting rule, numbers were averaged for each group and compared by ANOVA. In paraffin sections, 906 +/- 12 (SEM) neurons were counted, similar to previous calibrated data series, and results obtained from fixation with Methacarn or PFA were statistically indistinguishable. In celloidin sections, 912 +/- 28 neurons were counted-not statistically different from paraffin. In cryosections, 812 +/- 12 neurons were counted (underestimate of 10.4%) when fixed with PFA only, but 867 +/- 17 neurons were counted when fixed with PFA and GA. Vibratome sections had the most serious aberration with 729 +/- 31 neurons-a deficit of 20%. Thus, our analysis shows that PFA-fixed cryosections and vibratome sections result in a substantial numerical deficit. The addition of GA to the PFA fixative significantly improved counts in cryosections. These results may explain, in part, the significant numerical differences reported from different labs and should help investigators select optimal conditions for quantitative morphological studies.

Two distinct events, section compression and loss of particles (“lost caps”), contribute toz-axis distortion and bias in optical disector counting

Deformation of tissue sections in the z-axis can bias optical disector counting. When samples of particle densities are not representative for the entire tissue section, significant bias of estimated numbers can result. To assess the occurrence, prevalence, extent, sequence of events, and causes of z-axis distortion, the distribution of neuronal nucleoli in thick paraffin and vibratome sections was determined in chicken, rodent, and human brain tissues. When positions of neuronal nucleoli were measured in the z-axis, nucleoli were more frequent at the surfaces (bottom and top) of tissue sections than in the core. This nonlinear z-axis distribution was not lab-, equipment-, or investigator-specific, and was independent of age, fixation quality, coverslipping medium, or paraffin melting temperature, but in paraffin sections, was highly correlated with the tilt of the knife (cutting) angle. Manipulation of subsequent tissue processing steps revealed that two events contribute to z-axis distortion. Initially, a higher density of particles results at surfaces after sectioning, apparently due to section compression. Subsequently, particles can be lost to varying degrees from surfaces during floating or staining and dehydration, resulting in "lost caps." These results may explain different degrees of z-axis distortion between different types of sections and different labs, and reinforce the importance of checking z-axis distributions as a "quality control" prior to selection of guard zones in optical disector counting. Indirect approaches to assess section quality, such as resectioning in a perpendicular plane, yield additional artifacts, and should be replaced by a direct quantitative measurement of z-axis distribution of particles.

The use of stereological counting methods to assess immediate early gene immunoreactivity

The issue of whether profile and stereological counting methods are interchangeably accurate when assessing immediate early gene expression still needs to be resolved. To compare these two counting techniques, we quantified the expression of c-fos in the nucleus accumbens core and shell, and in the lateral septum as a control structure, of rats treated with neuroleptics. With the profile counting method, which relies on selective placement of a counting grid within a structure, we evaluated the density of c-fos labeled cells within a box of fixed dimension. With stereology, which applies random and systematic sampling methods, we used the optical fractionator method and counted the absolute number of c-fos labeled cells within the contours of each structure examined. Our results showed that the substantial increase in c-fos expression in the shell and core induced by haloperidol treatment was detected by both stereological and profile counting methods; in contrast, the weaker effect of clozapine on c-fos expression was detected differentially by the two methods. Whereas the profile counting method reported a reduction of c-fos in the core by clozapine, and an increase in c-fos in the lateral septum, these effects were not replicated using stereology. These findings suggest that stereological and profile counting methods do not always produce equivalent results. This may be particularly relevant when a measured effect is relatively small, and it is not distributed homogeneously within a structure. In this respect, the random and systematic sampling methods of stereology may yield more accurate and unbiased results than the profile counting method, and therefore may be preferred for a more accurate and thorough investigation of a treatment effect on immediate early gene expression in a specific brain region.

Sensitized Fos expression in subterritories of the rat medial prefrontal cortex and nucleus accumbens following amphetamine sensitization as revealed by stereology

Behavioral sensitization to the locomotor activating effects of amphetamine refers to the progressive, long lasting increase in locomotor activity that occurs with repeated injections. This phenomenon is thought to result from neuroadaptations occurring in the projection fields of mesocorticolimbic dopaminergic neurons. In the present study, we investigated the effects of amphetamine sensitization on Fos immunoreactivity (Fos-IR) in subterritories of the nucleus accumbens (core and shell) and medial prefrontal cortex (mPFC; dorsal and ventral) using stereology. Rats received five daily injections of amphetamine (1.5 mg/kg, i.p.) or saline. Behavioral sensitization was measured 48 h following the last injection, in response to a challenge injection of 1.5 mg/kg amphetamine. Sensitized rats showed a greater enhancement of locomotor activity upon drug challenge compared with their saline counterparts. Densities of Fos-positive nuclei were enhanced more in the dorsal than the ventral mPFC subterritory, whereas in the nucleus accumbens, densities of Fos-positive nuclei were increased more in the core than the shell of amphetamine-sensitized rats compared to controls. These results represent, to our knowledge, the first published report using stereological methods to quantify Fos-IR in the brain and suggest functional specialization of cortical and limbic regions in the expression of behavioral sensitization to amphetamine.

Celloidin mounting (embedding without infiltration) - a new, simple and reliable method for producing serial sections of high thickness through complete human brains and its application to stereological and immunohistochemical investigations

Celloidin mounting (embedding without infiltration) of the human central nervous system (CNS) proved to be superior to gelatin embedding for the production of serial sections ranging in thickness from 220 to 500 microm. After gallocyanin-staining, a comprehensive neuroanatomical as well as neuropathological survey of the human brain is possible, including diagnosis of Alzheimer's disease. Details of a fractionator analysis of the total striatal neuron number are described and the possible quantitative analysis of parallel immunohistochemically stained sections is discussed.

Stereological quantification of carboxyfluorescein-labeled rat lung metastasis: a new method for the assessment of natural killer cell activity and tumor adhesion in vivo and in situ

The function of natural killer (NK) cells is often studied by assessing in vitro levels of NK cell mediated lysis of target cells, or by assessing in vivo levels of lung tumor cell retention or metastatic colonization of intravenously injected tumor cells. However, these methods do not permit direct quantification and visualization of NK cells and their targets in vivo and in situ. Here, a new approach is described to visualize effector-to-target interactions as well as to estimate total numbers of targets in the lung, in vivo and in situ. MADB106 tumor cells were vitally labeled using carboxyfluorescein (CFSE) and intravenously (i.v.) injected into Fischer 344 rats (10(6) cells/rat). This mammary adenocarcinoma derived cell line is syngeneic to the inbred Fischer 344 rat and highly sensitive to NK cell activity in vivo. Effector-to-target interactions were visualized by immunostaining. Using the optical fractionator method, total numbers of CFSE-labeled MADB106 tumor cells were estimated in the left lung of the animals 5 min after tumor inoculation. To further demonstrate the usefulness of this approach in reflecting in vivo processes, rats were inoculated with MADB106 cells and simultaneously with a single i.v. bolus of either 1 microg/kg adrenaline or saline. Both lungs were removed 5 min later. Adrenaline caused a significant 80% reduction in the total number of lung CFSE-labeled MADB106 tumor cells, suggesting a rapid modulation of metastasis by stress hormones. This new approach facilitates the monitoring of effector-to-target interactions and the quantification of immune cell function or tumor adhesion in vivo and in situ.