New study uses ideas from string theory and quantum field theory to simplify calculations of transcendental numbers, like pi and Euler’s Zeta function.

Machine Learning models to aid faster and easier detection of brain cancer

Read time: 4 mins
22 Sep 2020
Machine Learning models to aid faster and easier detection of brain cancer

Representational image [Image Credits: Mindy Takamiya/Kyoto University iCeMS/ CC-BY-SA 4.0]

The human brain is around 90% of glial cells, which support the neurons or nerve cells and regulate the signals across them. Glioma is a fatal brain cancer resulting due to the abnormal growth of the glial cells. There are different variants or grades of glioma, depending on the microscopic traits and genetics of the cells or tissues involved. Determining the grade of glioma is vital to plan the course of treatment, which may include removing the tumour via surgery, or radiation therapy or chemotherapy, or a combination of both. Medical imaging techniques, like MRI (magnetic resonance imaging), is often used for detecting glioma. However, it is error-prone and time-consuming. A computerised system to classify tumours can address this shortcoming.

In a recent study, researchers at the Indian Institute of Technology (IIT) Roorkee, India, and Kyoto University, Japan, have designed a machine-learning algorithm to identify the grade of glioma with high accuracy. The study, published in the journal IEEE Access, was funded by the Ministry of Electronics and Information Technology (MeitY), Govt. of India.

"Machine learning-based models help radiologists and doctors in deciding with more precision," says Prof. Balasubramanian Raman from IIT-Roorkee, who led the study. "However, the primary issue in hospitals is the difficulty in generating MRI scans at different places, probably due to different MRI protocols. Moreover, there might be differences in MRI scans due to genotype differences in populations across the world," he adds.

Gliomas are graded on a scale of 1-4 depending on their location in the brain and the ability to spread to other parts. Grade 1 and 2 are slow-growing and are called low-grade glioma. Grade 3 and 4 spread rapidly to other parts of the brain and are referred to as high-grade glioma. They may recur even after a rigorous treatment.

Sometimes, distinguishing the two types is tricky as the cells involved in both the types look similar. Over time, a low-grade glioma can also turn into a high-grade one. There are also substantial differences in gliomas across patients as they vary in size, texture, and location in the brain. Therefore, accurate and rapid categorisation of glioma is vital for an effective treatment. MRI scans provide a wealth of information on the size, shape, and texture of the gliomas, which can help to generate a three-dimensional image of the scanned tumour. However, they need to be looked at by doctors, who may miss noticing features invisible to the naked eye.

Radiomics is an emerging field of medical oncology that hopes to address the problem of accuracy. It extracts many tumour-associated features from radiographic medical images like MRI, PET (positron-electron tomography), CT (computed tomography), and ultrasound, using computer software. By building models based on artificial intelligence, one can analyse these images better and diagnose and predict cancer with accuracy.

In the current study, the researchers have developed an artificial intelligence-based model that automatically classifies gliomas using radiomics. Artificial intelligence enables machines to take critical decisions of their own. They are fed with previous data to make them learn specific associations and then predict these associations with new data. More the data, accurate is this learning.

As acquiring MRI images is a lengthy process, the researchers used publicly-available MRI scans on BraTS (Brain Tumour Segmentation challenge) from 285 glioma patients. Of these, 210 had high-grade glioma while the rest had low-grade glioma. They trained their CGHF artificial intelligence-based model using this data and tested for its accuracy. They found that their algorithm surpassed several previous ones and was accurate about 98% of the time.

The proposed model also requires minimal human intervention. "The majority of the human-intervention is involved in training the model. Whenever the model predicts a new test case, experts might be required to verify the authenticity of results, especially when the chances of a tumour being a high-grade glioma are bleak", explains Dr Raman. The researchers are now considering testing their model with other recent and robust datasets like the Harvard MRI dataset. They also plan to work on the multi-class classification of graded gliomas for prediction of brain tumours that will throw more in-depth insight into a given glioma.

With India reporting about 28,000 new cases of brain cancer every year, using artificial intelligence to rapidly diagnose such conditions could help many patients.

"We believe that implementing this machine learning-based model in a clinical environment would be a significant improvement in the direction of automation. It will improve the efficacy of clinical processes with a reduction in human intervention," signs off Dr Raman.