High-throughput Image Analysis of Tumor Spheroids: A User-friendly Software Application to Measure the Size of Spheroids Automatically and Accurately

SpheroidSizer is designed to produce automated detection, delineation and measurement of the 3D spheroids, with remarkably reduced labor and acutely increased efficiency for large amounts of images. Figure 1A shows the workflow of SpheroidSizer. The core computation steps include automated initialization, active contour algorithm and contour quantification. After automated computation, the quality-control feature uses a combination of “Manual Initialize” and “Hand Draw” tools to salvage any imperfect segmentation. Figure 1B illustrates the detailed automated active contour algorithm. The initialization step (0th iteration) utilizes basic image processing steps to generate approximate size and location of the spheroid and generate a spherical initiation contour with an estimated size. The initiation contour feeds into the active contour algorithm. In turn it iterates to adjust according to the local image gradient and shape curvature. The active contour algorithm finishes when the contour stabilizes (converges), i.e. 477 iterations for this image, or when the pre-defined maximum number of iterations is executed. In this example, the initialization contour is purposely enlarged to better showcase the algorithm. In reality, the initialization is usually very close to the actual boundary and much fewer iterations are needed for the algorithm to converge. Subsequently, the algorithm takes morphometric measurements of the detected spheroid boundary. The major and minor axes of the spheroid are measured using MATLAB image processing toolbox (Figure 1C). The major axis is defined as the line segment connecting a single pair of the farthest points on the contour, which is referred to length (L). The minor axis is defined as the longest line perpendicular to the major axis, which is referred to width (W). In this case, the values of L and W are very close since the spheroid is spherical. The volume of the spheroid is calculated as V = 0.5 *L * W².

One of the features of SpheroidSizer is its automated detection of the boundary of the spheroids even on images with uneven or noisy background utilizing the active contour algorithm (Figures 2B-D). Computational processing of bright field images is often plagued by uneven background, which misleads adaptive thresholding-based methods to produce unwanted thresholding results. The issue is especially apparent when multi-well plates are used and walls of the wells may create shadowing effects on the images. However, because the active contour algorithm is not sensitive to the gradual shading change in background, it is able to identify spheroid segmentation in these bright field images with proper initialization. Figure 2 shows a few examples of images with uneven or noisy background, like uneven illumination (Figure 2B), debris (Figure 2C) or necrotic core (Figure 2D). With automated active contour algorithm, SpheroidSizer delineates these spheroids accurately in all these images as shown in the red contour at the lower panel of each figure.

The quality-control feature of SpheroidSizer is key to a high-throughput workflow. The “Manual Initialize” and “Hand Draw” tools are the valuable complimentary tools for this application. Among hundreds or thousands of images, it is unavoidable that the automated algorithm is not able to correctly detect the spheroids in some images. As illustrated in Figure 3A, when improper detection of the spheroid is caused due to the initialization step, i.e. improper size or location of the initiation contour in the image (top panel), the “Manual Initialize” tool works by allowing the user to properly define the location and size of the spheroid manually (lower panel). It triggers the active contour algorithm to initiate with manually defined contour and execute to converge on the desired outline. For these tough images like the original image in Figure 3B, the spheroid is located in a distracting and noisy background. SpheroidSizer is not able to properly identify the spheroid by the automated method (top panel) or by the “Manual Initialize” tool with proper initialization (middle panel). In this case, the “Hand Draw” tool can be used to manually draw the outline of the spheroid as illustrated in the lower panel. The program uses the user-defined boundary to measure the major and minor axes of the spheroid and calculate the volume. All the corrected results are immediately incorporated into the “Results Table” and can be accordingly exported.

To determine the performance of SpheroidSizer in larger data sets, we first compare operation time by analyzing the same set of 288 images using 1) manual measurements with microscope vendor supplied software; 2) SpheroidSizer with a single-core regular laptop; and 3) SpheroidSizer with a multi-core parallel computing performance workstation. The manual measurements follow our typical protocol prior to developing the software: the length and width of each spheroid are drawn by hand and measured using the vendor program (as seen the red lines in the top panel of Figure 4A); then the user copies down the values of the measurements. SpheroidSizer processes each image by generating the spheroid boundary (as shown the red outline in the lower panel of Figure 4A), measuring the axial major and minor length, and exporting the results in spreadsheets. As seen in Table 1, based on the calculation from 288 images, it takes an average of 31.67 sec to measure one spheroid per image manually; while it only takes SpheroidSizer less than 2 sec when running on a single-core regular laptop; and less than 1 sec when running on a 12-core performance workstation. Therefore, image analysis is over 18x faster per image using SpheroidSizer than manual measurements. It dramatically reduces labor when more than thousands of images are analyzed. Next, we compare the variability in measurements of the 24 spheroids shown in Figure 4A between manual measurements and SpheroidSizer. The 24 spheroids are measured three times by both methods; and the standard deviation of each individual spheroid is calculated. As seen in Figure 4B, the standard deviation from SpheroidSizer (green line and dots) is close to zero except for the three spheroids that are corrected at the quality-control step, which still show smaller standard deviation than those from the manual measurements method. All these indicate that SpheroidSizer performs the image analysis more efficiently and accurately.

We conducted a drug screen using human BON-1 3D tumor spheroids to find out which compounds in combination with an hsp90 inhibitor are the potential candidates for testing the anti-tumor effects in vivo. Human BON-1 3D tumor spheroids were grown on agarose-coated 96-well plates as described in the previous paper¹⁵. Eight different compounds with six serial dilutions plus media and vehicle were screened for their single and combinatorial effects with 10 nM and 20 nM hsp90 inhibitor in duplicates respectively. Two spheroids were used for each concentration of the individual compound or the combined compounds. Four 96-well plates with total 384 spheroids were used. All the spheroids were imaged at 0, 72, 144, 168, and 192 hr. A total of 1,920 images were produced from this experiment. It took SpheroidSizer only 30 min to complete the computational analysis of the 1,920 images with an additional 50 min for the quality control and data export. SpheroidSizer speeds up the image analysis process enormously. Figure 5A shows a screen shot of the folder arrangements and filenames for this experiment as an example for Protocol Step 3.3. Figures 5B-E shows the screen shots of the image analysis pop-up windows and results using SpheroidSizer as illustrations for Protocol Steps 4, 5, and 7. Taking the volumes of the 3D spheroids from the formatted results table exported from SpheroidSizer, we made graphs – growth of 3D tumor spheroids upon compound treatments versus time of treatments. Two representative graphs from this experiment are shown in Figure 5F and 5G. Figure 5F shows that the combined treatments of hsp90 inhibitor and cladribine (green line) inhibit the growth of 3D spheroids more than the single treatment of hsp90 inhibitor (purple line) or cladribine (orange line), suggesting that the combined treatments of hsp90 inhibitor and cladribine may have anti-tumor effects in vivo. Figure 5G shows that the combined treatments of hsp90 inhibitor and adriamycin (green line) does not inhibit the growth of 3D spheroids more than the single treatment of adriamycin (orange line) or hsp90 inhibitor (purple line), suggesting that the combined treatments of hsp90 inhibitor and adriamycin may not have anti-tumor effects in vivo. This experiment helped us better select the compounds to test their anti-tumor effects in vivo and SpheroidSizer is the key to the speedy experimental data analysis.

Table 1. Comparison of operation time on image analysis between manual measurements and SpheroidSizer when analyzing the same set of 288 images. Please click here to view a larger version of this table.

Figure 1. SpheroidSizer – an open-source software application for measuring the spheroid’s size. A) The core workflow of the application. B) Illustration of the active contour algorithm at different stages of iteration. Please note that the initialization contour (iteration 0) was purposely enlarged in order to showcase the algorithm. C) The major and minor axial length measurements and calculation of the volume by SpheroidSizer. L – major axis: the line segment connecting a single pair of farthest points on the contour (referred to length); W – minor axis: the longest line perpendicular to the major axis (referred to width).

Figure 2. Representative results from the automated segmentation of SpheroidSizer, showing robustness against various image conditions. A) Typical good quality images. B) Images with different brightness and contrast. C) Images with distracting debris. D) Images of spheroids with necrotic core. Images at the top panel of each figure are the source/original images; images at the lower panel of each figure are the quality-control images; and the red outline is the spheroid segmentation drawn by automated computation.­­

Figure 3. Illustration of the “Manual Initialize” and “Hand Draw” tools. A) The “Manual Initialize” tool allows the drawing of a fitting ellipse shape across the spheroid for initialization, when inaccurate spheroid segmentation occurs after the automated initialization. B) The “Hand Draw” tool allows accurate hand drawing of the spheroid boundary, when inaccurate spheroid segmentations occur with both automated and manual initialization. The blue line around the spheroid shows the initialization contour; the red outline is the identified spheroid boundary. Please note that the spheroid in “Manual Initialize” in A) and the spheroid in “Hand Draw” in B) is purposely enlarged to better showcase the tools.

Figure 4. Comparison of image analysis performance between SpheroidSizer and manual measurements when analyzing the same set of 24 images. A) Representative spheroids to show how the length and width of spheroids are determined by manual measurements and SpheroidSizer. Top 24 images contain hand-drawn length/width of each spheroid in red lines using manual measurements; lower 24 images (the same 24 images) contain computer-drawn spheroid boundary in red outline using SpheroidSizer. B) Standard deviation of length or width from three measurements on each individual spheroid.

Figure 5. A representative example of utilization of SpheroidSizer in drug screen – image analysis on the spheroids’ images that were collected from a drug screen using BON-1 3D tumor spheroids. A) A screen shot of the folder arrangements and filenames for this project. B) A screen shot of the advanced configurations window in SpheroidSizer. C) A screen shot of SpheroidSizer1.0 window with displayed Results Table. D) A screen shot of the format output file exported from SpheroidSizer. E) A screen shot of the list output file exported from SpheroidSizer. F) Growth of 3D tumor spheroids upon the treatments with hsp90 inhibitor and cladribine. G) Growth of 3D tumor spheroids upon the treatments with hsp90 inhibitor and adriamycin. Please click here to view a larger version of this figure.

Supplemental Files: SpheroidSizer software.  Please click here to download these files.