Applications

The following application have been developed on the basis of the Automated Cost / Result Optimization (ARCO) algorithm.

methods still under development:


Automated Algorithm and Parameter Selection (AAPS)

ARCO is based on online data analysis. For data analysis, in our case segmentation, we implemented an automated processing parameterization module for the LifeXplorer. It sweeps through a given parameter of the data analysis module and the parameter value with the highest amount of valid objects is then parameterized in the segmentation module. This approach ensures that one data analysis itself maximizes its outcome. Essential again is to define what this outcome quantitatively is before the experiment starts. Here the size of the cells has to be configured for an internal quality control after the segmentation.

 

AAPS

Figure 1 Experimental application of ARCO to optimize the number of cells detected. (a) Improvement of single nuclei detection. ARCO can be applied to maximize the nuclear detection, i.e. cell count, by selection from multiple segmentation algorithms, including global thresholding, supervised classification implemented in ilastik and CellProfiler, and a point source detector and changing their parameters if possible. (b) ilastik integrated in CellProfiler performed best and could improve sensitivity by 40% compared to a standard segmentation pipeline with global thresholding using ImageJ. (c) Example image HeLa cells labeled with nuclear dye Hoechst (1 μl). After pre-focusing with a 10x magnification, a 40x region of interest with the highest cell count is selected as the new image center. (d) For HeLa cells plated at a density of 50 to 10,000 cells per well, the nuclei count was increased up to 220%.

Application Specific Exposure Control (ASEC)

The first application of ARCO was to reduce phototoxicity in each optical channel. Based on Equation 1, the initial exposure time is chosen by the operator as the start value. The exposure time is automatically increased until saturation. The exposure time then is decreased incrementally and the segmentation efficiency is calculated each time, in our case to segment mitochondrial membrane potential or the Golgi complexes. Finally the minimal exposure time is taken where 90% segmentation efficiency is still possible.


LightExposure

Figure 2 Application Specific Exposure Control (ASEC). (a) Schematic illustrating the relation of light exposure to information content. As shown on the Y axis, different imaging functions need different levels of information content, i.e. segmentation works at lower information content whereas high-content feature extraction needs a high information content and thus higher light exposures. The operator, however, tends to choose the maximal dynamic range, which strongly decreases the survival time. (b) Schematic illustrating relation of information content and life time and phototoxicity. Classicallyintuitive operator light exposure configurations are used.  ARCO in contrast selects the optimal information content for a given data analysis (a), in our case a segmentation analysis, which requires lower light exposure in the dynamic range, and thereby minimized phototoxicity. 

Optimized Sampling Point Identification (OSAPI)

Within the autofocusing step, images are taken with a low 10x magnification. The optimized sampling point for the higher magnification is set to the center of the region of interest with the maximal number of relevant cells. To prove the method, we segmented nuclei using CellProfiler and ilastik that are plugged into the LifeXplorer control logic. The optimized sampling point identification lead to a doubled information density in the case of 10,000 cells plated and four- to sixfold enhancement for 5,000 cells.

LifeXplorer

Figure 3 Optimized Sampling Point  Identification (OSAPI). Intelligence Manager optimizing the XY-position to increase the cell count. After pre-focusing with a lower magnification, the region of interest is selected to maximize cell count. (b) Schematic illustrating the relation of light exposure to information content. As shown on the Y axis, different imaging functions require different levels of information content, i.e. segmentation operates at lower information content whereas high-content feature extraction requires higher signal-to-noise and thus increased light exposures. Classically, the operator, selects an exposure in the maximal dynamic range, which strongly decreases the survival time.

Refocusing Trigger (RFT)

The refocusing trigger can be derived as well from ARCO. Initially after the focusing step, the contrast value of each sampling point is saved in the LifeXplorer computing cloud. The RFT logic then triggers the microscope to refocus a position if the quality of the image of a sampling point drops under a certain threshold. Including ARCO, this logic can be generalized by adding a cost function, which leads to a refocusing if it is more efficient to refocus a sampling point then losing information content, because the focus has drifted. Refocusing a position is time consuming and thus creates costs in the dimension time and phototoxicity. The information content in the z-dimension can be quantified i.e. by a contrast value since a scientific question can only be answered if the quality of the image is suitable to extract relevant information. If the contrast value of an image of a certain sampling point drops under a predefined minimum quality threshold, the costs for losing information content have to be taken into account, i.e. because the optical spacial resolution is decreased. In our case a cost function is dynamically created on run time: The RFT logic simulated both behaviors every time, refocusing a position or not and calculated which behavior is more efficient based on then calculated ARCO efficiency. The phototoxicity could be reduced two to threefold by the refocusing trigger.


RFT overtime

Figure 4 Refocusing behavior over time using the refocusing trigger logic.


RFT architecture

Figure 5 Architecture of the refocusing trigger based on the LifeXplorer framework cuasing imaging acceleration and phototoxicity reduction

 

methods still under development:

Cutoff Frequency Detection (CFD)

CDF analyses the lowest sampling rate possible by applying ARCO. Measuring an optical channel in the case of microscopy, the sampling rate configured by the operator is taken first. CDF then simulates lower frequency sampling rates and detects the lowest sampling rate where it is still possible to detect 99% of the cells correctly. To analyze mitochondrial energetics a high sampling rate is necessary. We performed experiments in HeLa cell loaded with TMRM (50 nM), imaging with the Olympus IX81 at 20x magnification for a period of 8 hours, at 30 second increments. For the Hoechst channel the LIFE X-PLRORER used nuclei tracking based on the method of least squares. The highest sampling frequency set to 30 s and lowest was configured to 20min. After 20 min, on runtime, the LifeExplorer then was able to detect a suitable sampling frequency. The lowest sampling rate of 20min was still suitable to track 99% of the objects. Using the cut-off frequency detection (CFD), the amount of data could be reduced to 51% which is almost equal to a twofold information density increase. The computation time to segment the nuclei was reduced tenfold.

 

Model Based Autofocus Algorithm Selection (MBAAS)

MBAAS

Figure 6 Surface modeling and acceleration workflow. Automated selection of the best focus algorithm and acceleration of the autofocus step. Part (a) shows the workflow how to accelerate the focusing step while the maximal amount of in-focus images as reached. Sampling a reference well of the slide with many positions, a surface model can be created. The model serves as a ground truth reference which can be used to automatically select the best focus algorithm. In a next step, the optimal sub-sampling for all others well can be computed, which maximizes the focusing while the maximal amount of in-focus images is kept. (b) shows a result of the surface modeling using a third order fit. (c) visualizes exemplarily three different algorithms with the same accuracy, but difference precisions. Part (d) shows an acceleration curve depending on the number of focused and approximated spots.

 

Region Of Interest Based Alignment (ROIBA)

Once a live cell sample has to be taken down from the microscope in order to do a washing etc. like it was necessary for the reassembly of the Golgi apparati, the position of the cells after putting the sample back will not be exactly the same due to mechanical impreciseness of the sample carriers. We therefore implemented a re-alignment function. After putting the sample back to the microscope, by a macro a bigger area (5x5) is taken as a stitched image. The original reference image before taken away the sample is then searched and the positions are realigned by the difference vector.