New Delhi: Researchers at IIT-Madras (IIT-M) have developed algorithms that could help detect multiple sclerosis (MS) which, since it is visible as several small lesions, could be easily missed.
MS is a disease in which the protective sheath covering the nerves gets destroyed, disrupting the communication between the brain and the rest of the body. This leads to difficulty in speech, sight and the ability to move.
"The task of accurate delineation of regions (segmentation) of the brain affected with MS is a difficult and time-consuming affair. Owing to this, significant variability can be observed in the regions marked by different radiologists on the same image. In case of MS, only 50 percent of the marked area would match each other," Ganapathy Krishnamurthi, professor in IIT-M's department of engineering design, who led the research, told IANS.
He added that the team's research focusses on development of automated methods to perform accurate segmentation of disorders such as MS and glioma.
Explaining further, he said that these segmentations were important for doctors to obtain quantitative metrics for treatment monitoring and planning, as well as for surgical operations.
The number of multiple sclerosis patients has increased in India in recent years. It is estimated that there are between 100,000 and 200,000 MS patients in India. According to the All India Institute of Medical Sciences, which carried out a study in 2013 on the patients of multiple sclerosis it treats, about 70-80 percent of patients were in the 18-35 age group.
Krishnamurthi shared that while working in collaboration with Thiruvananthapuram's Sree Chitra Thirunal Institute of Medical Sciences and Technology, the team identified that accurate labelling of disorder-affected regions in brain MRI could be a difficult affair due to its "complex shape and vague boundaries".
"Moreover, it is a tedious task since radiologists cannot visualise in 3D and the task needs to be performed slice by slice," he said.
He added that this led to research on automated methods for identification of glioma (brain tumors) affected regions from MRI images.
"However, the core algorithms developed in the process were such that they could be used in the detection of other disorders as well. Multiple Sclerosis is a chronic disease which is visible as several small lesions which can be easily missed. This being a particularly difficult task, we decided to extend the research scope and tackle this problem as well," he said.
The symptoms of MS include weakness or numbness of limbs, blurring, partial or complete loss of vision, slurred speech, dizziness, tremors, lack of coordination and tingling sensation or pain in the body.
The team, comprising Suthirth Vaidya and Abhijith Chunduru, final year integrated masters (B.Tech+M.Tech) students from the engineering design department under the guidance of Krishnamurthi and M. Ramanathan, used technology known as 'Deep Learning', which is inspired by advances in neuroscience and is loosely based on the interpretation of information processing and communication within the nervous system.
"Deep Learning are recent methods in machine learning developed based on the interpretation of how human brain and nervous systems work - neural networks. These networks consist of stacked layers consisting of several mathematical models of neurons, which is the computational equivalent of information processing in the brain. Although these methods have been around for more than a decade, recent developments in computational resources have made large and complex networks with near-human performance possible," Krishnamurthi explained.
Voice recognition on Android smartphone, Google's self-driving car and automatic photo tagging feature on Facebook are all powered by Deep Learning.
The team, which emerged victorious in the recently held Longitudinal Multiple Sclerosis Segmentation challenge at International Symposium on Biomedical Imaging (ISBI) 2015, New York, is currently in the process of building a software tool that can be used by clinicians.
"Our next steps in this endeavor would be to test extensively with more clinical data to assess the effectiveness of the software and subsequently deploy the software for use by our clinical collaborators. Based on the performance in a clinical setting (purely for evaluation) we will try to get regulatory approval for our software. Since training accurate models require large amounts of data, ethical committee approvals from various hospitals would be required. We are already in collaboration with Sree Chitra Thirunal Hospital and are confident of seeing the product put in use in a span of two to four years," he said.
So, will it make MS treatment/diagnosis cheaper?
"These methods when implemented can substantially reduce the time and cost for diagnosis of various brain diseases like MS. The algorithms for image analysis are basically a tool for diagnosis and aids clinicians to judge progression of disease and efficacy of therapy. For instance, in large clinical trials these automated algorithms can be used to analyze patient data," Krishnamurthi added.