Lennert S. Ploeger

Development of 3D chromatin texture analysis using confocal laser scanning microscopy

Introduction: Analysis of nuclear texture features as a measure of nuclear chromatin changes has been proven to be useful when measured on thin (5–6 μm) tissue sections using conventional 2D bright field microscopy. The drawback of this approach is that most nuclei are not intact because of those thin sections. Confocal laser scanning microscopy (CLSM) allows measurements of texture in 3D reconstructed nuclei. The aim of this study was to develop 3D texture features that quantitatively describe changes in chromatin architecture associated with malignancy using CLSM images. Methods: Thirty-five features thoughtfully chosen from 4 categories of 3D texture features (discrete texture features, Markovian features, fractal features, grey value distribution features) were selected and tested for invariance properties (rotation and scaling) using artificial images with a known grey value distribution. The discriminative power of the 3D texture features was tested on artificially constructed benign and malignant 3D nuclei with increasing nucleolar size and advancing chromatin margination towards the periphery of the nucleus. As a clinical proof of principle, the discriminative power of the texture features was assessed on 10 benign and 10 malignant human prostate nuclei, evaluating also whether there was more texture information in 3D whole nuclei compared to a single 2D plane from the middle of the nucleus. Results: All texture features showed the expected invariance properties. Almost all features were sensitive to variations in the nucleolar size and to the degree of margination of chromatin. Fourteen texture features from different categories had high discriminative power for separating the benign and malignant nuclei. The discrete texture features performed less than expected. There was more information on nuclear texture in 3D than in 2D. Conclusion: A set of 35 3D nuclear texture features was used successfully to assess nuclear chromatin patterns in 3D images obtained by confocal laser scanning microscopy, and as a proof of principle we showed that these features may be clinically useful for analysis of prostate neoplasia.

Implementation of Accurate and Fast DNA Cytometry by Confocal Microscopy in 3D

Background: DNA cytometry is a powerful method for measuring genomic instability. Standard approaches that measure DNA content of isolated cells may induce selection bias and do not allow interpretation of genomic instability in the context of the tissue. Confocal Laser Scanning Microscopy (CLSM) provides the opportunity to perform 3D DNA content measurements on intact cells in thick histological sections. Because the technique is technically challenging and time consuming, only a small number of usually manually selected nuclei were analyzed in different studies, not allowing wide clinical evaluation. The aim of this study was to describe the conditions for accurate and fast 3D CLSM cytometry with a minimum of user interaction to arrive at sufficient throughput for pilot clinical applications. Methods: Nuclear DNA was stained in 14 μm thick tissue sections of normal liver and adrenal stained with either YOYO-1 iodide or TO-PRO-3 iodide. Different pre-treatment strategies were evaluated: boiling in citrate buffer (pH 6.0) followed by RNase application for 1 or 18 hours, or hydrolysis. The image stacks obtained with CLSM at microscope magnifications of ×40 or ×100 were analyzed off-line using in-house developed software for semi-automated 3D fluorescence quantitation. To avoid sectioned nuclei, the top and bottom of the stacks were identified from ZX and YZ projections. As a measure of histogram quality, the coefficient of variation (CV) of the diploid peak was assessed. Results: The lowest CV (10.3%) was achieved with a protocol without boiling, with 1 hour RNase treatment and TO-PRO-3 iodide staining, and a final image recording at ×60 or ×100 magnifications. A sample size of 300 nuclei was generally achievable. By filtering the set of automatically segmented nuclei based on volume, size and shape, followed by interactive removal of the few remaining faulty objects, a single measurement was completely analyzed in approximately 3 hours. Conclusions: The described methodology allows to obtain a largely unbiased sample of nuclei in thick tissue sections using 3D DNA cytometry by confocal laser scanning microscopy within an acceptable time frame for pilot clinical applications, and with a CV small enough to resolve smaller near diploid stemlines. This provides a suitable method for 3D DNA ploidy assessment of selected rare cells based on morphologic characteristics and of clinical samples that are too small to prepare adequate cell suspensions.

Confocal 3D DNA Cytometry: Assessment of Required Coefficient of Variation by Computer Simulation

Background: Confocal Laser Scanning Microscopy (CLSM) provides the opportunity to perform 3D DNA content measurements on intact cells in thick histological sections. So far, sample size has been limited by the time consuming nature of the technology. Since the power of DNA histograms to resolve different stemlines depends on both the sample size and the coefficient of variation (CV) of histogram peaks, interpretation of 3D CLSM DNA histograms might be hampered by both a small sample size and a large CV. The aim of this study was to analyze the required CV for 3D CLSM DNA histograms given a realistic sample size. Methods: By computer simulation, virtual histograms were composed for sample sizes of 20000, 10000, 5000, 1000, and 273 cells and CVs of 30, 25, 20, 15, 10 and 5%. By visual inspection, the histogram quality with respect to resolution of G0/1 and G2/M peaks of a diploid stemline was assessed. Results: As expected, the interpretability of DNA histograms deteriorated with decreasing sample sizes and higher CVs. For CVs of 15% and lower, a clearly bimodal peak pattern with well distinguishable G0/1 and G2/M peaks were still seen at a sample size of 273 cells, which is our current average sample size with 3D CLSM DNA cytometry. Conclusions: For unambiguous interpretation of DNA histograms obtained using 3D CLSM, a CV of at most 15% is tolerable at currently achievable sample sizes. To resolve smaller near diploid stemlines, a CV of 10% or better should be aimed at. With currently available 3D imaging technology, this CV is achievable.

A restaining method to restore faded fluorescence in tissue specimens for quantitative confocal microscopy

The aim of this study was to develop a procedure to remove the TO‐PRO‐3 fluorescent dye from tissue sections and restain with TO‐PRO‐3, still allowing calculation of DNA content and distribution by confocal laser scanning microscopy (CLSM). This would allow repeated measurements on the same tissue sections and prevents loss of tissue material from valuable clinical samples. Thick sections (14 μm) were cut from a paraffin block of adrenal tissue and stained using TO‐PRO‐3. Image stacks were acquired by CLSM. Thereafter, three destaining approaches were tested based on incubation, at different temperatures and durations, in the medium that is normally used to dissolve TO‐PRO‐3. The same areas were imaged again to measure residual fluorescence and were subsequently restained and imaged again. The intensity of the images acquired after initial staining and restaining were compared. A number of 3‐D (texture) features computed after segmentation of nuclei were compared as well. The best destaining result was obtained by incubation of sections at 37°C in preheated medium twice for 20 min. On average, the 3‐D feature values were comparable with those after initial staining. With the described protocol it is possible to remove TO‐PRO‐3 fluorescence from tissue sections that can successfully be restained with minimal influence on fluorescence intensity and nuclear chromatin distribution.

Racial Differences in 3-D Nuclear Chromatin Patterns of Prostate Cancer

Purpose: There is a significant difference in prostate cancer incidence and stage corrected mortality between African-American (AA) and Caucasian-American (CA) men. These differences have largely been contributed to social- economic factors, yet variation in prostate cancer related gene expression has been found as well. The aim of this study was to analyze whether these differences are reflected also in the 3-D distribution patterns of the nuclear chromatin. Materials and Methods: Prostatectomy sections from 21 prostate cancer patients (10 AA and 11 CA) were cut and nuclear DNA was stained with TO-PRO-3. 3-D image stacks of selected malignant areas were obtained by confocal laser scan- ning microscopy. Image analysis was performed using in-house developed software for 3-D semi-automated segmentation and computation of DNA content and our previously developed 3-D nuclear texture features. The power of these features to discriminate between AA and CA patients was established by univariate ROC and linear discriminant analyses, stratify- ing for prognosis. Results: Five 3-D texture features discriminated between AA and CA men irrespective of prognosis, 27 features had dis- criminative value for AA and CA men in the subgroup of bad prognosis patients, and 8 features in the good prognosis subgroup. Several features had additional discriminative value in multivariate discriminant analysis. Conclusions: There are differences in the 3-D nuclear chromatin distribution between AA and CA men with similar prog- nosis. This is further evidence that the differences of prostate cancer in AA and CA men are not only related to socioeco- nomic differences, but also to genomic differences.

Computer science meets medical science.

In the past 20–30 years at least two technical advances have made quantification in pathology at a large scale possible. At the time the computing power of affordable computers increased, interest to deploy tasks emerged that were previously very laborious to perform. Combined with digital camera systems, many new possibilities became available. Examples of such measurements that were previously very tedious if not impossible to perform are to measure the optical density of a nucleus, the area of a cell and the ratio of the red versus the green signal. Together, these techniques became known as “Quantitative Pathology” and were pioneered by Baak et al. and others [1–3,7]. Often, papers that report these novel techniques, especially papers on texture features [8], are accompanied by an impressive amount of mathematical formulas in the materials and methods section. This also goes for the paper of Nielsen et al. in the present issue of Cellular Oncology [14]. To the reader, it may be a daunting task to delve into the details and to get complete understanding of the measurements being performed. However, to the technically inclined (as the author of this editorial, having a degree in computer science) these formulas are essential to be able to reproduce the measurements and to optimize one’s own technique for a certain task. In general, a single formula is able to communicate a much larger part of source code that actually implements the described technique. As is true for a natural language (e.g., English) the language of mathematics is a common language which is understood by a large audience. From this point of view the current practice with many formulas is the most concise way to present this information. However, in an interdisciplinary field, where computer science is applied in medical science, these technical sections might shy away readers that are less used to the language of mathematics. As a worst-case consequence, the entire paper can be put aside while the novel technique might have a large impact on patient diagnosis or prognosis. Mostly in-text formulas are still in use [6,12,13], but sometimes they are moved to the appendix section [11]. The use of an appendix for this purpose is a good compromise between the ability to be able to reproduce an experiment and to keep the paper readable for a larger audience. The criterion for this decision should be whether the total number of formulas is large enough (∼5 or more) to justify an appendix. Another solution is to present the techniques in a table in the appendix and provide an exhaustive list of references [9]. The advantage is that the paper is very orderly, but a disadvantage might be that it is more complicated to repeat the measurement in a different institute. When science develops, a basic set of primitives is obtained and usually it is sufficient to refer to these common concepts without providing the details. This concept is known as abstraction, and is used in everyday life. Eventually, this will reduce the need to present the level of detail often seen in current publication. However, new technologies in medicine will always require custom-made solutions and hence it is expected that difficult papers will always be part of medical literature. As an example, focus is currently moving from 2D applications [5] towards 3D applications [4,15,16]. An important driving force is the development of better imaging techniques in confocal laser scanning microscopy like the development of 4π microscopy [10]. This technology will present detail that was previously imperceptible. Understanding the corrections required to achieve this resolution is still a daunting task, but researchers that are aiming at a better image quality should be informed about all important aspects. Another subject of study is improvement of image quality by deconvolution [17]. Deconvolution by itself is a very complicated technique, and it is an important subject of several conferences. Therefore, it is anticipated that literature on deconvolution, with an abundance of formulas, will soon appear in medical journals. In conclusion, technical papers form an important part of medical literature and this will also be the case