Grade 11 students build AI tech to detect oral cancer

Every year 4th Feb is observed as World Cancer Day with the primary goal to significantly reduce illness and death caused by cancer. The new World Cancer Day 2022-2024 campaign theme, Close the care gap, is about identifying and addressing the barriers that exist for many people around the world in accessing the care they need.

On this occasion, we thought it is best to highlight the work of two young students, studying in grade 11 in Mumbai, who have developed an AI-based solution to make oral cancer detection accessible and affordable in India.

Kudos to you both Maanav & Aditya! What is your innovation all about and what problem does it solve?

Oral cancer, the third most prevalent cancer in India, is easily treatable if diagnosed at an early stage. In rural India, 60% of oral cancer cases are diagnosed in advanced stages. The solution for this is mass early-stage screening, which is not easy to implement as it requires experienced doctors and healthcare personnel in large numbers.

We have designed a device that scans the entire oral cavity for potentially cancerous lesions without professional intervention. Our Invention - the MouthScope - uses Artificial Intelligence (AI) to make oral cancer screening more accessible by automating it.

Mouthscope works on the principle of autofluorescence: it emits low-wavelength light which excites fluorophores in normal mucosa and not cancerous tissues, causing cancerous tissues to have a darker color. The phone camera captures this and sends it to the machine learning model which predicts with 86% accuracy.

By using a smartphone and classifying images through machine learning, Mouthscope eliminates the need for extensive infrastructure for mass screening of oral cancer, making it more attainable for rural India.

What was the problem with the existing techniques that made you come up with this innovation?

In India, there is a mismatch between people susceptible to oral cancer (the ones who fall into risk categories like tobacco consumers and smokers) and infrastructure available for diagnosis to them, especially in rural areas: close to 80% of tobacco consumers live in rural India. The gold standard to diagnosis is biopsy which needs trained professionals and laboratory facilities. Moreover, the process of biopsy is an invasive procedure. Hence we need screening protocols and devices for such susceptible patients where only the patients with suspicious and potentially cancerous lesions are shortlisted from the susceptible population ad subjected to further diagnostic tests.

Screening is a practical solution and we have several devices available as well.

However, the screening devices (Velscope, Vizilite, OralID) available in the market have one or many of the following shortcomings

• extremely expensive
• not accurate
• not automated
• not AI-based

Our device “Mouthscope” aims to solve this by eliminating the need for expensive diagnostic tools, trained personnel. Requiring only one affordable device (target price set at ₹15,000) for mass screening of such susceptible patients (target accuracy set at 80%) in a non-invasive manner.

The Mouthscope also addresses the problem of follow-ups. Follow-ups are essential after treatment, as the recurrence rate of squamous cell carcinoma is 32.7%. However, there is a lacking medical infrastructure, especially in rural areas, to facilitate the recommended number of follow-ups. Being a device that can attach directly to a smartphone, the Mouthscope can also be used for follow-ups without requiring much infrastructure.

Affordability and accessibility were the key features that we envisioned for Mouthscope.

How long did it take for you to conceptualize the idea?

We both connected on this topic during the lockdown in April 2021.

Maanav “I saw my dentist parents unable to attend to their patients during the lockdown and that got me thinking"

Aditya “our entire family was in shock when an unsuspected mouth ulcer was diagnosed as an advanced stage of oral cancer for my grandfather."

We then decided that we need a device that can help in oral screening especially suspicious and potentially cancerous lesions for earlier diagnosis.

How long did it take to build the prototype? What was the process?

We took about 7 months to build the present prototype with 6 different designs evolving and making it accurate at every stage. Right from getting the correct light source to procuring the correct lens to generate autofluorescence, using different cameras for image capturing, and then programming the accurate AI model for image processing.

Can you describe the features of your product/ technology in simple terms?

The Mouthscope involves two components:

• a hardware component (a physical device with short-wavelength light and filters), and
• a software component (computer vision and web-app).

The hardware component consists of a 3-D printed attachment for a mobile phone. The attachment houses a dichroic long-pass filter placed directly on top of the phone’s camera as well as a light source of wavelength 405 nm attached in a circular arrangement around the lens. The light from this source is directed towards the mouth. Inside the mouth, due to the short wavelength of the light, fluorophores in the normal mucosa are excited and emit light of different wavelengths than the cancerous or potentially cancerous lesions. On re-entering the device, this light is also incident on the dichroic long-pass filter, which only allows light above 450 nm to pass through, resulting in potentially malignant lesions having darker color than the normal oral mucosa. This light is directed to the mobile phone’s camera.

How can patients make use of Mouthscope?

We have also designed a web app that allows patients to view the YOLOv5 model’s predictions in real-time, allowing a focus on self-diagnosis. The web-app inputs information about the patient. The device stores the images and uploads them to the cloud, where the resnet_v2 model makes a binary prediction (classifying as cancerous or non-cancerous). These images and the patient’s information are presented to a doctor, who verifies the predictions. If the predictions are incorrect, the model is retrained based on this, and (unless it fails an accuracy test) the model is updated.

How big is your team? What were their roles and contributions?

We are a team of two that is Aditya and Maanav but were mentored and supported by many professionals from different fields. The technical support and lab support were from OMOTECH lab, Mumbai.

We were guided and mentored by Ritu Jain, Dr. Deepa Nair, Oncologist, Tata Memorial Centre; Dr. Vivek Borse, DST INSPIRE Faculty, Centre for Nanotechnology; Edmund Optics; Dr. Jigna Pathak, Professor MGM Dental College, and Hospital; Dr. Tejas Mhatre, Oral Surgeon; Dr.Arjun Singh, Research Fellow at Tata Memorial; Dr. Rajeev Chitguppi, Periodontist & Dental researcher; Mr. Sharan from Mantle Labs.

Did you conduct any study with this? What were the results?

Yes.

We tested our device with 4 potentially cancer patients to see the efficacy of our solution. The four patients had previously been tested with early squamous cell carcinoma, erythroplakia, and 2 cases of oral submucous fibrosis. They were tested positive by biopsy for cancer.

With our device, we were able to see a visible color difference between the normal tissue and the potentially cancerous lesions in the patients’ mouths. As expected, we observed that in the case of the normal oral mucosa, the image captured by the device had a bright green hue while the potentially cancerous lesion did not undergo autofluorescence and was visibly darker. The machine learning models we trained were also able to detect these lesions.

Our neural networks had an accuracy of 86%, and in the future, we hope to capture more images with our device and increase our dataset size. This will allow us to not only improve precision but also carry out multi-class classification in place of the current binary classification, resulting in more detailed screening.

Did you present your work anywhere? How was the response?

We have presented our work and research at various competitions like the IIT Tech Fest, IRIS national science fair, where we won the GRAND AWARD in the biomedical engineering category and will now represent India at. The Regeneron International Science and Engineering Fair (ISEF) is to be held in Atlanta, Georgia, USA. At ISEF, finalists will compete with 1800+ participants from over 80 countries, regions, and territories around the world.

Have you published your work in any journal?

Not as yet, we are working on it.

What are your further plans with this innovation?

• Our vision is to empower rural India to fight against oral cancer.
• Our immediate goal is to deploy Mouthscope in at least 10 villages in rural India in 2022.
• We are in discussion with subject experts, oncologists, technology institutions to help us strengthen our product.
• We have to get the final product ready for production we also understand that it requires ICMR Clearance.

About the innovators:

Maanav Kothari is currently studying in grade 11 at Dhirubhai Ambani International School for IBDP. He is extremely passionate about robotics and has participated in various national and international robotics competitions. He is also a part of the school FRC team and enjoys CAD and designing devices and cars and bringing them to life with technology.

Aditya Mehta is currently studying in grade 11 at Dhirubhai Ambani International School and has a keen interest in Programming and Artificial Intelligence. He believes that we need to leverage technology to solve and simplify the challenges our country faces - MOUTHSCOPE is one such initiative.

Tags: