Express Healthcare
Home  »  Interviews  »  Deep learning is ruling the roost in healthcare

Deep learning is ruling the roost in healthcare

0 273
Read Article

Institute of Advanced Study in Science and Technology (IASST), Guwahati, an autonomous institute of the Department of Science & Technology, Govt of India, has developed an artificial intelligence (AI) based algorithm as an aid to rapid diagnosis and prediction of oral squamous cell carcinoma. Dr Lipi B Mahanta, Associate Professor, CCNS, Institute of Advanced Studies in Science and Technology (IASST) in interaction with Sanjiv Das, elaborates on the research activity undertaken by her and her team, the technique followed and the challenges faced

IASST has developed an artificial intelligence-based computer diagnosis framework for accurate oral cancer diagnosis. Tell us more about the technique.

We have used a technique called deep learning. In artificial intelligence, medical image analysis has been done since the 1950s. First medical image analysis was done with sequential application of low-level pixel processing and mathematical modelling. Next, supervised techniques under a machine learning paradigm arrived, which is still very popular. We have also developed algorithms using this paradigm, for binary classification, that is whether a given image is normal (non-cancerous) or abnormal (cancerous). We have achieved 100 per cent accuracy in a few algorithms in this study.

Currently, deep learning is ruling the roost. Here, algorithms models (networks) are composed of many layers that transform input data (e.g. images) to outputs (e.g. disease present/absent) while learning increasingly higher-level features. The most successful type of models for image analysis to date is convolutional neural networks (CNNs). CNN’s contain many layers that transform their input with convolution filters of a small extent. These filters or layers can learn low-level features (like straight edges, simple colours, and curves) on their own. Input images are recognised, or classified, by slicing them into smaller blocks and then matching the low-level features in them. For this study, we have done multi-class classification, that is we have classified the images into different stages of oral squamous cell carcinoma, a diagnosis which is most vital for prognosis. We have used two separate approaches to the study.

First, there are many pre-trained deep CNNs already available. We have implemented four of those networks, which we deemed most suitable for our application, namely Alexnet, VGG-16, VGG-19 and Resnet-50, and then tweaking the final layers specific to our dataset. Secondly, we have built a CNN network from scratch specific to our dataset. The accuracy of our NEW CNN is 97.5 per cent as compared to the accuracy of 92.15 per cent of Resnet-50 (the best of the four implemented pre-trained networks). The details are available in our paper.

How do you plan to conduct field trials?

We will conduct a field trial through our primary customers. Our primary customers are all personnel who are involved with the collection and diagnosis of cervix cancer using pap smears. These include:

  1. Every hospital across India, large or small, public or private; every health service centre in districts and villages.

  2. Also, NGOs (like ASMI, based in Guwahati) who are actively working in awareness and collection of relevant cancer samples through screening camps.

We will collaborate or chalk out an MoU with all these customers, enabling them to use our software free for one month. They will give feedback on the same, based on which we will improve the algorithm or the interface if needed.

How do you plan to reach out to the masses?

We plan to host the software through the Internet. This means that the software need not be purchased by any user. It will be kept in a web server, supported by IASST. Access to it will be permitted to the registered users, for a fee. Using a central server will entail that whenever modifications/upgradations are needed in the software it can be done easily and globally. We will reach out to the mass through our primary customers.

What benefits did you receive from the Department of Science & Technology in this regard?

This study was a part of our regular scientific activity. We have achieved all this with the usual infrastructure. But if we want to proceed and achieve what we dream, we will need hand-holding.

How many scientists took part in this project? Any challenges you faced while working in this project?

This study was initiated and led by me since the last five years. Over this period some collaborations were struck for exchange of technical and clinical advice, and collection of the database. Four scholars/lab assistants in this team carried out the experiments and data acquisition work, namely Tabbasum Yesmin Rahman, Elima Hussain, Navarun Das and Himakshi Borah. Institutions collaborating with us were IIT Kharagpur, B Borooah Cancer Institute (BBCI), Ayursundra Pathological Lab, Arya Wellness Centre and Gauhati Medical College and Hospital (GMCH). Dr Chandan Chakraborty of IIT Kharagpur extended his technical expertise. Doctors involved in the study for clinical advice as well as the creation of the dataset were Dr JD Sharma (BBCI), Dr Anup Das (Ayursundra Path Lab and Arya Wellness Centre), Dr Chandana Ray Das, Dr Manjua Chakravarty and Dr Ratna Kanta Talukdar of GMCH.

There were immense challenges faced till now. I should not spell out the red tape problem a scientist faces in every corner in such interdisciplinary and multi-institutional studies. Ethical Clearances (for Human Studies) had to be taken, MoUs were signed, and relevant, helpful and interested doctors were mined out.

Data acquisition strategy was another main challenge. Not only to create a well-oiled mechanism for the flow of patient-to-images but also the standardisation of the images acquired. Deep learning requires high-end hardware and software support. The team is still struggling for that.

Do you plan to conduct any more research in cancer care with the help of IT?

Definitely. I am leading a group of around 10 students who have worked or are currently working in several domains of a cancer diagnosis. Our work in cervical cancer is commendable. We are simultaneously working on breast cancer, lung cancer, stomach cancer and medulloblastoma (which is a special type of brain cancer seen especially in children). Our approach is that we adopt those types which are most relevant in our part of the country and are recommended for study and software development by clinicians.

Leave A Reply

Your email address will not be published.

Attend an exclusive webinar to know more about biomarkers in diagnosis & therapy management of heart failure.
Register Now
close-image
 Introducing Smart Autoinjector: Changing the paradigm of usability, cost & size
Know More?
close-image
Want to know more about "Omic's" and it's relevance?
Register Now
close-image