Automated aid for screening of oral cavity lesions – A feasibility studyTctd@1232021-12-01T09:03:57+00:00
Automated aid for screening of oral cavity lesions – A feasibility study (Ongoing) (Ongoing)
PI: Prof. Amit Sethi, Electrical Engineering Department
Tata Fellow: Akanksha Shreshtha (2019-21)
Although oral cancer is preventable when detected early, it has the highest mortality count among cancers prevalent in India. Fortunately, early detection of oral cancer can easily be done by mass screening using pre-screening visual inspection of the oral cavity, followed by a screening biopsy of suspicious cases. Such mass pre-screening is infeasible in India due to the lack of trained manpower. Additionally, a large number of false positives in order to reduce life threatening false negatives will lead to overburdening of oral pathologists who need to do the biopsy and screening diagnosis thereon. Taking primary healthcare (PHC) workers out of the field for months-long training to pre-screen potential candidates for biopsy is also infeasible. An accurate but mobile and (semi) automated pre-screening solution to analyse images of the oral lesions along with other indicators such as tobacco usage can be beneficial for designing and executing a mass screening program.
We propose to determine the feasibility of developing a smartphone app powered by artificial intelligence (AI) to flag candidates for oral biopsy on which a primary healthcare worker can be trained in a few hours. We zeroed in on the idea of a smartphone app due to the ready availability of smartphones with cameras and LED lighting to take images of the oral cavity. Using an app with a computationally light AI-based inference engine or by transmitting images to a heavy-duty AI inference server over the mobile network, will be the backbone of a subsequent affordable and scalable solution. In this work, we propose to determine the feasibility of developing the AI-based decision engine whose key determinant is the size of the image corpus required for training a machine learning decision engine, with acceptable false positives and very low false negative rates.