Providing quality medical services and surveying the rural population in a country as big as India is extremely complex:
- 67 percent of India’s population lives in rural areas;
- 90 percent of medical facilities are located in cities;
- And for a million people in the country, there are only three radiologists.
Health education programs were aware of the importance of medical screening throughout the country — this, in turn, increased the workload for the few available radiologists who work at their limits. As a result, radiologists have even less time to make detailed diagnoses. In such conditions, artificial intelligence (AI) can be a real breakthrough, which Indian radiologists need so much.
Integration of deep neural networks
The MIRIAD project, led by Associate Professor Debdut Shitom, is investigating methods for implementing deep neural networks (DNN) to increase the effectiveness of radiological screening using AI — this will significantly improve the quality of healthcare throughout India.
“The challenge,” — says Debdut, “is to cope with a wide variety of different medical images, including X-rays, computed tomography (CT), magnetic resonance imaging (MRI) and histopathological examination of whole slides (WSI). There are also domain features in the character of the image data, as well as the modality and specificity of some of them. For example, the number of channels in these images is not always limited to one (shades of gray) or three (RGB palette). These factors complicate the problem of object detection, ” — says Debdut.
Despite these problems, the first major achievement of the MIRIAD project was the development of a deep neural compression mechanism for mammography. The team took a huge step forward by conducting a comparative analysis of existing DNNs for mammography and chest X—ray extensions.
Inspired by the technology of deep image compression, the Debdut team developed a convolutional model (similar to autoencoder) for relevant compression mammograms while maintaining characteristics. The use of arithmetic coding for working with a large number of spatial redundancy allows packaging code with higher density – this leads to a shorter length of one information bit.
The Debbut team used Intel® AI DevCloud, which runs on Intel® Xeon® Platinum 8160 and Intel® Xeon® Gold 6128, to train deep neural networks (GNS) using Intel® Math Kernel Library (Intel® MKL) and the Intel® Math Kernel Library for GNS (Intel® MKL — DNN) and Intel® Distribution for Python version 3.5.
Intel engineers helped the team achieve maximum optimization by speeding up DNN training in terms of variable accuracy. Access to a variety of resources has been a critical factor for success.
For developers starting similar projects with Project MIRIAD, the Getting Started with Intel AI DevCloud manual is a good introduction to optimized Intel software tools.
Architectures and methods for teaching AI to compress radiological images and effectively screening radiological images will significantly improve the speed and accuracy of diagnosis. Developers have many great features and high — tech tools — now they can use AI to support manufacturers of medical imaging equipment and computer-aided detection and diagnostics (CADx).
Editor of IMD News