Image Processing - Geometric Illusion and 3-D Estimates


Examining blood smears or urine samples through automated microscopy has umpteen number of challenges, some of which were addressed in Bharath’s article on Challenges and Opportunities in Automated Microscopy. In this blog, I would like to continue by outlining many more challenges associated with automated microscopy.

While examining urine samples or blood smears through a microscope, the cell morphologies doesn’t exactly concur with the textbook definition, especially, when you have cells occluding each other. As a human, you can use your intuition to conclude if a platelet is occluding a RBC, but in automated microscopy you have to “teach” your algorithm to “learn” such occlusions. Not only do these complexities increase the confusion between classes for a computer, but many-a-times they make decisions tough for a human too.

Let’s take an example, below are a set of images of a malaria infected RBC and a normal RBC with a platelet on top of it. These images look similar and you might find it a bit tough to differentiate between them, but, what’s herculean is that you would have to build an algorithm to decipher such occlusions.

This is just one example from a plethora of permutations explaining confusion between different types of cells.

Let’s assume you have solved the above challenge. But, there are tougher challenges that are yet to be dealt with. For example, in the microscopic examination of urine, the sample contains multiple layers, hence the cells could be in multiple orientations. Again, as a human one could use intuition to identify the cells but we have to take “sufficient and variable” data to teach the algorithm most if not all of these orientations. Shown below are images of RBC in different orientations from a urine sample when examined under a microscope. Similar problems exists for other objects of interests such as casts, crystals, WBCs etc


To add-on to the above problems, the blood report should contain accurate estimates of volumetric parameters such as Mean Corpuscular Volume (MCV), Mean Corpuscular Hemoglobin (MCH), etc. While the estimation of these metrics seems quite trivial, they are much complex than what meets the eye. Let me take you through an analogy; Assume that you have the top view of a shopping mall and you are tasked to estimate the average height of people in the mall just using the top-view. That’s a seemingly simple problem to define and a real tough nut to crack, isn’t it? Yes, computing volumetric or area based parameters of RBCs in a blood smear is a similar problem with several layers of complexities brought in by the requirements to ensure that the computation is not only precise and accurate, but, also quick!