Cervical cancer kills nearly 73,000 women every year in India, many of them from rural areas. Screening for cervical cancer, which takes around ten years to fully develop, can help detect the cancer in its incipiency and therefore prevent it from becoming dangerously malignant. In a Pap smear test, which is usually used for cervical cancer screening, cells are taken from the uterine tract and studied by cytologists. Cancerous cells look nothing like a normal cell. They often have multiple nuclei, are abnormally elongated and can be visually identified by trained lab experts. But this test requires a lot of resources – laboratories, equipment, facilities for the transport of specimens and, importantly, trained people who can visually detect cancerous cells from the smears. The test is also very expensive, which limits its reach.
To wipe out cervical cancer in a developing country like ours, there is a need for cost-effective technology for large-scale screening. In the Department of Chemical Engineering at IIT Madras, there is a project underway which could provide just that. Professors Sridharakumar Narasimhan and Basavaraja M. Gurappa are working on an automated algorithm for the screening of cancer cells. By doing away with the necessity of having cytologists search for and find cancerous cells in Pap smear tests, computer algorithms can simplify the process of screening for cervical cancer, making the test affordable and accessible for large numbers of women across the social strata. As Prof. Sridhar puts it, if the visual identification of cancer cells could be automated, a cancer screening test wouldn’t need more than a janitor and some equipment.
For such automation, we need a computer to perform the role that a cytologist’s eye and brain play in spotting cancerous cells in a Pap smear sample. More precisely, we need “computer vision,” a field that involves the study and development of computer methods for acquiring, processing, analyzing and understanding images. Cancer cells are elongated in shape, and can be approximated mathematically as ellipses, which can then be detected by computer algorithms.
Suyog Sawala, a B.Tech from the Department of Chemical Engineering at IIT-M who graduated in 2013, worked along with Prof. Sridhar and Prof. Basa on this integral part of the problem – the detection of elliptical shapes by computer vision – as his final-year project. This involves lengthy and complex computations, and is a challenging problem in computer vision. To complicate matters, the real world is far from ideal and elliptical objects are never perfect – they come in altered shapes and in forms that are clumped together, or “occluded.” Suyog’s work focuses on improving existing methods of detecting elliptical objects in two-dimensional images speedily and efficiently, and is highly important for Prof. Sridhar’s work on cancer detection.
For his project, Suyog was recognised with the “Best Project” award in the Mechanical Engineering division at the 2013 Jed-i Project Challenge, which is an annual event aimed at identifying and showcasing the best final-year engineering projects, held at the Indian Institute of Science, Bangalore. Members of the jury at the event comprised experienced industrialists and researchers from different disciplines, while participants showcased a whole spectrum of projects from prosthetic arms to soil analyzers to bio-energy production from waste.
Delving into the details of Suyog’s work, Prof. Sridhar says that the geometrical principles underlying basic shape detection through computer vision are simple and intuitive. Shapes are detected using techniques like “Canny edge detection” and the “Hough transform.” He explains: “Suppose you are looking for a line in an image. First, you would identify all the pixels on the edges of different shapes in the images. Edge detection by a computer is based on differences in the brightness of image components. For example, in the image of a blood smear, the sharp contrast between the cell and the surrounding fluid can be used to identify the edges of the cell. Every pixel on the edges is given a geometric position, an x-coordinate and a y-coordinate. To detect lines, the computer would then need to identify collinear edge pixels – pixels which lie in a straight line.”
He continues: “A line can however, can be defined completely by only two points. This is not true when identifying ellipses, as a minimum of five parameters, namely, the axes lengths – major and minor, the orientation of the ellipse, the position of the centre of the ellipse – its x- and y-coordinates, are required to uniquely identify it. Thus, the problem of identifying ellipses is computationally intensive while being sensitive to the aspect ratio of the ellipse. Here, the Hough transform technique comes to the rescue. Imagine any two edge pixels and assume they are the end-points of the major axis of some ellipse. Then consider any other edge pixel to be lying on this ellipse. These three pixels together define a unique ellipse. Now count the number of object pixels through which this ellipse passes. If this count is greater than the threshold, our assumed ellipse is actually an ellipse and is thus successfully differentiated from the rest of the pixels.
However, these generalized versions of the Hough transform are in a “high-dimensional” space – space that has more dimensions than the usual three. Working with such spaces can consume a lot of a computer time and memory. There are other challenges too.
Says Suyog, “Detecting shapes when a clear outline is not visible, detecting individual objects in a complex cluster-forming pattern, detecting them faster, making the algorithm robust to noise, while also being able to adapt to detect slight variation in shapes are part of the challenge.” Suyog, along with Professors Sridhar and Basa, came up with new algorithms that are not only computationally less demanding but are also effective in the presence of noise or occlusion in the image. Their algorithms are also quite successful at detecting slightly altered shapes – for example, the sickle-shaped blood cells that are symptomatic of sickle cell anemia.
So why was a project involving image processing and cancer screening taken up in the Department of Chemical Engineering? Suyog answers: “The project was initially started because of its major applications in colloidal sciences – shape detection algorithms could be used to quantify the number density, size and orientation of colloidal particles from microscopy images. As the project reached its primary milestone, its applications in other areas were explored to realize its full potential. Cancer cell detection was one such area.”
Although Suyog has now graduated from IIT-M and is working as a professional with Shell, he is simultaneously working with his project advisors towards publishing a paper on his project work, and on finding feasible ways to commercialize this project. If Prof. Sridhar and Prof. Basa find ways to successfully apply Suyog’s work in cervical cancer detection, it will contribute towards making the screening tests cheaper and more accessible. And if that happens, it could make the lives of millions of Indian women a little brighter.
(With inputs from Professors Basavaraja M. Gurappa and Sridharakumar Narasimhan, Suyog Sawala and Adarsh Chavakula, a B.Tech third-year student of Chemical Engineering and a member of the team working on cancer detection.)