What do Lucky Charms, bone marrow and cord blood stem cells have in common?
We can use image analysis to evaluate each mixed population. While your marshmallow to cereal ratio may define you as the leader of your breakfast club, unleashing the imaging data within your preclinical research or clinical trial may indeed save lives. We know – we help our clients do it every day.
Our St. Patrick’s Day Image Analysis
Our goal was to develop an automated algorithm to analyze the number and distribution of marshmallows and toasted oat pieces within a box Lucky Charms breakfast cereal, and to evaluate the mixed population relative to where a sample was taken from the cereal box. In other words, to maximize marshmallow content in your bowl, should you aim to pour from the top, middle, or bottom of the box?
From top to bottom of box
| Marshmallow Density
by Count (%)
by Area (%)
Original Image and Image Overlay (from top of box to bottom)
HOW DID WE DO IT?
Think of Lucky Charms as, like bone marrow or cord blood, a large heterogeneous population of cells wherein different cell types exist in varying amounts. The marshmallows and the toasted oat pieces represent different cell types. To answer our “existential” breakfast question, we divided the cereal box into five roughly equal bowls, ranging from Bowl 1 (top) through Bowl 5 (bottom). Each bowl was spread out as a single uniform layer and imaged with a standard high-resolution SLR camera. Once each image was acquired, our analysis algorithm visually clarified the edges between tightly packed pieces and defined morphological bounds of size and shape to include in the analysis set. A pair of masks was then generated for the pieces identified as marshmallows vs. cereal, utilizing additional filters and color channel operations. Each segmented piece is individually numbered and exported along with the associated measured area to a spreadsheet. The algorithm then aggregates the count and area data, and reports the sum total for each type of piece in a given bowl. Marshmallow content is then characterized both in terms of count density, the proportion of pieces identified as marshmallows, and area density, which compares the 2D space occupied by marshmallows in the image to the overall bowl composition. From there, the algorithm outputs a pseudo-colored image overlay of marshmallows and toasted oat pieces colored green and gold (respectively), co-registered with the image of the original bowl. After all, what good is quantitative data without a visual audit trail?