Fig. 1: In a blood smear (top), blood cells have three typical types of distribution (bottom): sparse (left), good (center) and clumped (right). Areas of the three types are illustrated in the middle row in blue, green and red, respectively.
Reproduced, in part from, Ref. 1 © 2010 IEEE
The microscopic analysis of blood smears is a routine procedure in medical diagnosis. Now, Wei Xiong at the A*STAR Institute for Infocomm Research in Singapore and collaborators have developed and tested an algorithm for the automated initial classification of smear image areas in the context of high-throughput screening.
To prepare a blood smear, a drop of blood is simply placed on a glass microscope slide and then spread using a wedge. The sample is then chemically stained so that infected blood cells can be identified and counted.
This technique, however, creates many areas to examine on the slide. Also, smear thickness varies between slides, depending on the size of the blood drop, the angle of the spreader, and the speed of spreading. Moreover, at one end of the slide, where the smear is thickest, blood cells tend to clump together, while at the thin end of the smear the cells are generally sparse and unevenly distributed (Fig. 1).
It is surprisingly difficult and time-consuming for manual cell enumeration and detailed diagnosis using all the areas in the smear. Manual selection of ‘good working areas’ where the cells are evenly distributed, well separated and form a thin layer of cells is subjective and inconsistent. “The purpose of automatic working area selection is to pick out from the smear as many good areas as possible that are suitable for subsequent detailed computer analysis,” explains Xiong.
The team’s algorithm rapidly and objectively categorizes good, clumped and sparse working areas based on key features, such as the sizes and distribution of objects that are visible on different parts of the slide and the numbers of cells that these objects contain. This allows the reliable identification of good working areas using relatively little computational power. “This is a great improvement over manual image assessment, which is a tedious job prone to human error,” says Xiong.
To test the algorithm, the researchers took more than 15,000 images of malaria-infected blood smears using a digital camera linked to a motorized microscope. The method proved to be highly accurate, achieving a hit rate of nearly 90% for all areas in the validation set of 140 images, and robust, with a hit rate of nearly 80% for a test set of nearly 5,000 images.
The researchers hope that application of their algorithm will greatly improve accuracy and reproducibility of smear image analysis, reducing labor and greatly increasing the reliability of high-throughput microscope screening of blood smears.
The A*STAR-affiliated researchers mentioned in this highlight are from the Institute for Infocomm Research.
