DNA Ploidy

Such abnormalities were originally implicated in cancer development over 100 years ago by German biologist Theodor Boveri, although by the year 2013 it had become clear that cellular aneuploidy is a driving force in carcinogenesis. At present, the most widely used approach to detecting aneuploidy is DNA cytometry — a well-known research assay that involves staining of DNA in the nuclei of cells from a tissue sample, followed by analysis using quantitative flow cytometry or microscopic imaging.

ICGI has utilised DNA ploidy in both clinical diagnostics and research since its beginning and carried out close to 50 000 ploidy analysis on tumour samples. Ever since we moved from Flow cytometry to Image Cytometry in the early ‘90s in the Department of Pathology, we have developed our own DNA ploidy systems, for commercialisation as well as for our own use. However, cytometry has a limitation; the tissue must be disintegrated to make suspensions of pure nuclei, and this requires a certain amount of material. Hence, when analysing biopsies from, e.g., early prostate cancer, such as during active surveillance, successful measurements have only been achieved in about 75% of the biopsies. For this reason, it would be very advantages to be able to do ploidy analysis directly in routine histological sections. A section-based method would also simplify the clinical use and make the method much more available in routine pathology.

With the ongoing digitalisation in pathology, there is also a demand to replace our advanced motorised microscopes with routine scanners and do the ploidy analysis directly in WSIs. We have very promising preliminary data for both adaptions, but so far, the results are inferior to the established method based on high-resolution automated microscopy and monolayers of nuclei.

DNA ploidy in sections

Measurements in 5um thin sections in prostate tissue result in the same rough classification (diploid vs non-diploid) as measurements in monolayers from the same block in 88% to 93% of the cases. The resolution of the IOD histograms is clearly inferior in section measurements, leading to a higher Coefficient of Variation (CV), resulting in both false positive and false negative observations.

There are several possible sources of error in the section measurements, mostly related to the fact that the nuclei are not whole, but cut profiles, errors that will affect small and large nuclei (profiles) differently:

• Nuclear segmentation
• Volume modelling
• Glare and diffraction
• Cell type classification
• Texture Analysis

Nuclear segmentation, which is straight forward in a monolayer is an incredibly challenging task in sections, will be addressed with CNNs (see Technology > Segmentation > Nuclei segmentation on page 127). Volume modelling, i.e., estimating the DNA content in a whole nucleus based on measures of a cut profile is also a challenge. This is partly because nuclei in several types of tissues are far from spherical and partly because we do not know where the profile in an assumed spherical nucleus is cut from. We have parallel materials with 5um and 12um thin sections (the latter mostly containing whole nuclei) where we obtain data that might allow us to develop better volume modelling or provide a better understanding of the impact of this source of error. Glare and diffraction are optical phenomena that affect measurements in the microscope, and have different impacts on objects of differing size. It should be possible to devise algorithms that, at least in part, may correct these errors. All tumours contain several different cell types that contribute differently to the measurement results. We will create a learning set of cell-type nuclei (see Technology > Segmentation > Tumour tissue segmentation, page 126) and implement a deep learning-based algorithm to classify nuclei according to cell types, and hence improve the accuracy between internal control nuclei (that is used to establish the diploid position in the histogram) and the nuclei of interest (in most cases nuclei from epithelial tumour cells).

It seems reasonable to expect that we, through partial success on several of these attempts, will be able to raise the ploidy accuracy per nucleus above 90%. As an overwhelming majority of non-diploid tumours have more than 25% polyploid or aneuploid cells, this goal should secure a clinically safe method for DNA ploidy assessment from tissue sections.

DNA Ploidy from whole slide imaging

We have developed software for grabbing and classifying nuclei imaged with commercial scanners. Currently, the correlation with respect to ploidy classification of monolayers imaged with microscope- and scanner-based methods is around 90-95%.

Even here we find that the histograms CV’s are increased (although far from what was reported for sections above) causing errors in ploidy classification. The correlation varies with the type of aneuploidy, as the typical classification error is found for hypo/hyper-diploid tumours and 5c exceeding-rate (5cER) tumours, that are erroneously classified as diploid. Aneuploidy in the form of hypo/hyper-diploidy is a frequent finding in gynaecological cancers, but rare in, e.g., colorectal cancers, hence the correlation depends on the type of tumour measured.

Commercially available scanners are made for visualisation of colour images, primarily of tissue sections. They use lenses with lower numerical aperture, typically based on a standard 20x lens (and adding another 2x lens for “40x” magnification), typically also have a smaller total magnification as well as pixel resolution compared to images obtained with a light microscope equipped with a 40x lens. Although the images produced by scanners are non-inferior to microscopes when compared for diagnostic accuracy in pathology, the resolution is inferior and evident when performing objective image analysis. Another challenge with most scanners is the lack of flexibility when it comes to the light source. Whereas microscopes typically use a range of filters to optimise the light source for a given task (such as a 546nm green filter for the Feulgen technique) this option is normally not available in scanners.

To improve ploidy classification based on WSIs, one basically must improve the scanners. We are currently following two routes to improve image resolution in scanned images. One route is based on a collaboration with the Chinese company MBM where we work together to improve the image quality and resolution and to implement a 546nm filter in front of the light source. We expect a system with our specifications implemented for delivery early in 2021, and as we receive this scanner free of charge, we have agreed to attempt to validate the use of this for DNA ploidy through a series of analysis of patient cohorts with known ploidy distributions.

Our second route is HighRes, the development of a high-resolution scanner optimised for image analysis. For both routes, we will need to revisit the cell classification algorithm used to ensure that this is optimal for scanner images. So far, we have been using a technical adaption, allowing us to retrain the algorithm using scanner images of the same nuclei used for training the microscopy algorithm. In principle, this should be fine as long as the optimal training for a scanner is not required to be more extensive than for microscopy. The alternative will be to increase the training set by manually correcting current classification results from the scanner and retraining the algorithm. A separate software, Scan-to-Nuc, has been developed to grab the nuclei in the WSI. The following tasks will be verified or approved:

• Scan-to-Nuc; segmentation
• Cell classification (rule file)
• Effect of green filter
• Effect of reduced resolution (comparing 20x and 40x using the microscope platform)
• Dark current, glare and diffraction

Ploidy Work Station (PWS) is commercially available through Room4 Group Limited, England.