By: Ganapathy Krishnamurthy
The research: Developing algorithms that separate the images of tumours or lesions in the brain from the background. The algorithms will be developed using artificial intelligence, and will help in improving diagnosis and treatment of glioma and multiple sclerosis.
We conduct research on medical imaging techniques and analysis, using images from MRI and CT scans. The research group’s focus is on the images of the brain and its diagnosis of disorders such as glioma and multiple sclerosis.
Our primary research is to develop algorithms that separate the images of tumors or lesions from the background. This will help machines identify lesions and troubled spots in the images and delineate them from the rest of the image.
Today, the identification and delineation is mostly being done manually. It is a tedious process and has scope for errors. Automated methods learnt from previously diagnosed images identify abnormalities with greater accuracy and uniformity. This leads to improved identification methods and thus better diagnosis and more targeted medical interventions.
- Story of an extraordinary death: How MRI machines work, and can (in rare cases) kill
- Mumbai MRI Death: Three Chandigarh hospitals order safety review
- MRI machine death: Staff who accompanied deceased not from radiology department, were untrained
- Smartphones may help search for cosmic rays from black holes, supernovae: Study
- Bi-annual MRIs helps high risk breast cancer patients: Study
- An MRI that may detect Alzheimer’s
Dual degree students of the department, Suthirth Vaidya and Abhijith Chunduru, joined the research group, guided by Dr Ramanathan Muthuganapathy and me. They started their research focusing on developing algorithms using a technique in artificial intelligence called ‘deep learning’. It is an advanced branch of study called ‘neural networks’ that seeks to enable machines to simulate human beings in recognising pictures and sounds through interpretations of how our brain processes them.
For example, human beings can easily identify a person in different photographs. To enable a machine to do so, however, very complex algorithms need to be developed. Once these algorithms are in place, the accuracy in recognising the images are much better and faster.
The research team applied the ‘deep learning’ techniques to ‘segment’ lesions of multiple sclerosis, a complex brain disorder. The lesions of multiple sclerosis are very small and appear in complex shapes with vague boundaries on MR images. The algorithm was able to mark out the lesions at an accuracy equivalent to human performance.
Evaluating the exact performance of such algorithms on medical images is a difficult task, since radiologists themselves mark the affected regions with high variability. The algorithm developed in the research lab currently performs at an accuracy equivalent to this inter rater performance.
The algorithm developed by our team was recently adjudged as the best performing one among 34 entries, including ones from Harvard University and University College London, at a competition, part of an IEEE International Symposium of Biomedical Imaging, held in New York.
Author is the Assistant Professor, Department of Engineering Design, IIT Madras.