Overcoming Autofluorescence Variability: Innovative Strategies for Accurate Data

In this article, Senior Image Analysis Scientist Caitlin Rutherford shares her thoughts on the challenge of managing autofluorescence in multiplex immunofluorescence (mIF) studies. Autofluorescence, while a widespread natural phenomenon, can significantly interfere with the accuracy of mIF analysis by overlapping with true positive signal.

Caitlin explores various techniques, including using artificial intelligence and deep learning classifiers, to effectively combat autofluorescence at both the tissue and cellular levels, ensuring more precise and reliable results in your research.?

The Autofluorescence ‘Problem’?

Multiplex Immunofluorescence analysis has been a significant breakthrough in the thorough interpretation of complex biological systems. It simultaneously enables the colocalisation of multiple markers in the same section by utilising different fluorophore dyes (with different wavelengths) to bind to target proteins/molecules. This allows for a deep investigation of complicated biology within a single tissue section through the analysis of multiple markers simultaneously, so a deeper investigation into biological processes and interactions is possible?– for example, into specific complex phenotypes and co-expression spatial information.

One of the many challenges in multiplex analysis is the presence of autofluorescence, especially when the signal overlaps with true positive staining in other channels. Autofluorescence is where biological structures (e.g. red blood cells/collagen) naturally emit a wavelength of light that can interfere with the applied fluorophores corresponding to specific markers of interest. The interaction of these natural structures with the mIF-specific dyes can create an overlap in the signal. As expected, this creates challenges when aiming to accurately detect only true positive staining (while avoiding any overlap of autofluorescence false positivity).

Figure 1 – The example above clearly shows autofluorescence (pink) is “bleeding through” to a specific marker channel (green). This highlights how challenging it can be to analyse images with autofluorescence overlapping true positive signals – examples of this overlap are highlighted in yellow.

Autofluorescence is a natural process that can be introduced at the tissue processing stage. Many different methods should be considered before the analysis stage to mitigate the effects of autofluorescence.

For example, autofluorescence can be limited by choosing fluorophores with longer wavelengths and well-separated, longer excitation wavelengths. Blue and green fluorophores (DAPI/ Alexa Fluor 405 and FITC/ Alexa Fluor 488, respectively) have a closer wavelength to the natural emission wavelength of autofluorescence. In contrast, red and far red fluorophores (Alexa Fluor 647, 750 and Cy5) are further away, so these should be chosen where possible.

Chemical quenching agents (such as Copper Sulphate, Sudan Black B and Glycine) can also be used to reduce autofluorescence chemically. Each agent has advantages for specific tissues, e.g., Copper Sulphate, which is particularly useful for formalin-fixed tissue. The fixation method used is also an important consideration as these can affect autofluorescence - formalin fixation increases autofluorescence, whereas ethanol/ methanol fixation decreases autofluorescence. Shorter fixation periods can also aid in limiting autofluorescence.

In addition to these approaches, autofluorescence blockers can block light in a specific spectral range, reducing the visibility of autofluorescence and increasing the visibility of true fluorophore signals. The autofluorescence blocker chosen will depend on the tissue used; for example, trypan blue - which absorbs ~580-620nm - is best used on collagen-rich tissues, liver, kidney, respiratory and neural tissues.

These are just a few examples of the methods and techniques that can limit autofluorescence before image analysis. However, measures can be taken to combat autofluorescence during image analysis, and it is important to consider the best analysis strategy before developing algorithms to ensure the most accurate results are generated.

?

How Does OracleBio Deal with Autofluorescence?

Our image analysis scientists use Visiopharm and Indica Labs HALO? to analyse images. We have developed strategies to combat large and small areas of autofluorescence across both image analysis softwares. The figure below (Figure 2) shows both large and small areas of autofluorescence on a single mIF image, highlighting the challenging nature of analysing mIF images, particularly with variable autofluorescence.

Figure 2 – Example within one image of (A) large areas of autofluorescence (green) and (B) smaller areas of autofluorescence (yellow). This figure highlights how variable autofluorescence can appear across a study image set and why the approach must be considered before developing analysis.

Image Credit: INCISE.

Large areas of autofluorescence (Figure 2A above) often occur in large biological structures in the tissue, such as vessels containing red blood cells (RBCs) or collagen.

In both Indica Labs HALO and Visiopharm, deep learning (AI) classifiers can be developed to recognise and remove these areas from the final area for analysis. These deep learning classifiers are created by annotating representative markup areas for each class to ensure that all variety is captured across the markups. It is important to create markups on areas with enough context to teach the neural network exactly what constitutes autofluorescence compared to what is not.

Before training the classifier, the resolution can be changed to ensure enough context is visible to the neural network during training. Similarly, the minimum size can be used to ensure that the classifier detects areas of a specific size. So, for both of these parameters, if large areas of autofluorescence are removed with a deep learning classifier, the resolution can be relatively low as little detail is needed. In contrast, a relatively high minimum object size can be chosen to pull out large areas & avoid fragmented detection.

Before training, specific channels can be selected to detect true autofluorescence vs. true positive staining. Ensuring that the autofluorescence channel is selected when training the classifier is important to provide full training on the appearance of autofluorescence in the study.

Deep learning classifiers can detect and remove autofluorescence in a variety of ways. Some studies will succeed with only two classes (tissue and autofluorescence), where any detected autofluorescence can be removed from the analysis area by converting to negative annotation in the tissue layer – Figure 3 (below).

Figure 3 – The example above shows how to combat large areas of autofluorescence (A). A two-class classifier can be created to detect tissue shown in blue vs autofluorescence, as shown in green (B). The mask output generated by the classifier (B) can be converted into annotations (C) in the tissue annotation layer & inverted to ensure that the classified areas of autofluorescence are removed from the area for analysis.

Image Credit: INCISE.

Other studies will benefit more from incorporating autofluorescence into a classifier with multiple classes (e.g. tumour, stroma, white space, autofluorescence) – Figure 4 (below).

As this type of classifier has the regions of interest (ROIs) incorporated in the classifier and autofluorescence, there is no need for any conversion to negative annotations in the tissue layer. Both of these classifiers can be created and utilised in Visiopharm and HALO.

Figure 4 – The example above shows how a multi-class classifier can be created to segment autofluorescence and regions of interest, such as tumour and stroma. In this example, the above classifier would be incorporated into further analysis algorithms and only the regions of interest (tumour/ stroma) would proceed for analysis.

Image Credit: INCISE.

Both methods of removing autofluorescence by classifiers have their advantages and disadvantages and should be considered prior to development. A two-class classifier (Figure 3) is typically easier to create, but large amounts of negative annotations on an image can slow down image panning or processing. Whereas, a multi-class classifier (Figure 4), can be harder to create, especially if there is challenging morphology across the image set. However, removing autofluorescence from the classifier level - rather than negative annotations - is visually cleaner and won’t slow down image processing or loading times.

As well as large areas of autofluorescence, small areas of autofluorescence (e.g. individual RBCs through the tissue) can also create challenges when analysing mIF images, especially if they are mistaken for true positive signals. Small areas of autofluorescence are shown in Figures 1 and 2B above.

Within both softwares, there is also the ability to remove small areas of autofluorescence from within the cell analysis algorithms. As with markers of interest, thresholds can be selected for the autofluorescence channel within the cell analysis algorithm to detect autofluorescence above a selected threshold as positive. This autofluorescence threshold can be incorporated into phenotypes to ensure that anything detected as a true phenotype is also negative for autofluorescence. This means that any phenotype with an autofluorescence signal above a specific (autofluorescence positive) threshold will be discounted from the final phenotype population as it is likely actually autofluorescence and not a truly positive cell.

While highly effective, creating analysis settings with the aim of only quantifying phenotypes that are negative for autofluorescence also has challenges. It can be difficult to choose an effective threshold across multiple images, especially if there is varying autofluorescence intensity across the images. Mean intensities can offer insight into how variable the image set is and whether this approach would be feasible for the study.

As detailed above, a deep learning classifier can be developed to combat smaller areas of autofluorescence by utilising a larger resolution and smaller minimum object size. This can be particularly effective when using a two-class classifier, as shown in Example 1 above. However, depending on whether many negative annotations are present, this may reduce processing and loading efficiency across the images.

Another technique to remove AF from images during the cell analysis stage is to subtract the AF signal directly from other channels affected by AF. This can be performed in Visiopharm, where the process creates a new feature image that effectively removes areas of overlapping AF signal but leaves a specific target signal, allowing for cleaner downstream target thresholding.

Figure 5 (below) highlights an area inside the tumour that is rich in CD8 immune cells [cyan, 480nm] but also contains prominent red blood cells (RBCs) autofluorescing at the same wavelength. When thresholding or CD8 positivity using the CD8 channel, CD8 negative cells whose boundary overlaps RBCs can be misclassified as CD8 positive due to the AF signal present in the RBCs. However, using the feature image for thresholding allows for more accurate detection of true CD8 cells.

Figure 5 - (A) Area inside Tumor showing CD8 immune cells [cyan, 480nm] and also red blood cells (RBCs) autofluorescing at the same wavelength, DAPI in blue; (B) CD8 channel only; (C) Autofluorescence [AF] channel only; (D) a new feature created in VIS subtracting the AF channel from CD8 channel (E) thresholding for positive CD8 cells [yellow overlay] using the raw CD8 channel or (F) the AF channel feature.

In Summary

While multiplex immunofluorescence provides valuable insights into complex biological systems, autofluorescence poses a considerable challenge.

I have outlined practical strategies to mitigate autofluorescence throughout the mIF workflow, focusing on advanced image analysis techniques. You can effectively distinguish true positive signals from autofluorescence by leveraging deep learning classifiers and precise thresholding, ensuring more accurate and meaningful study outcomes.?

As imaging technology and AI-driven analysis tools continue to evolve, I anticipate even more refined approaches to overcoming the autofluorescence 'problem', paving the way for increasingly precise biological insights.?

If you’d like to learn more about anything covered in this article or to get support with your studies, please don’t hesitate to get in touch and arrange a short call with our experts!