Report on research undertaken in the year 2024-25:

In the past year, we developed AI tools for oral cancer screening that can be embedded within a mobile phone to classify white light images of the oral cavity captured from smartphone cameras into ‘suspicious’ and ‘non-suspicious’ categories based on the presence of pre-cancerous or cancerous lesions, thereby developing a point of care, AI-assisted, phone-based diagnostic device, ensuring availability, accessibility and affordability. From the rigorous experimental studies conducted in the previous year, on the architecture of the model, data cleaning, data augmentation, handling class imbalances, loss functions, and training strategies, we had learnt that the MobileViT-v2 model trained with random horizontal-vertical flip and rotations with weighted crossentropy loss gave the best performance in terms of sensitivity and specificity on the test data. With this as the baseline, we then moved to testing the trained model on the field (external validation) data. We observed a domain shift between the training and the field data, which led to a dip in the performance of the model. Additionally, the field images had EXIF metadata flags such as ‘Orientation’, which were not present in the training data, affecting the way how images were read and processed. We resolved these issues by incorporating EXIF-based augmentations and white-balancing in the training and preprocessing phases, respectively, which improved the performance of the model on the field data. In the next phase of integrating the trained and traced AI model into mobile applications, we noticed that several functionalities that were available on the computer platform were not available on the mobile platform. We had to modify the processes on the computer platform as per the mobile platform to ensure a matched performance between the phone and the computer. With this, we were able to achieve good performance metrics on not only the source data but also the field data, across all platforms.

Report on contributions to CNI activities:

I was involved in content generation using AI for Random Process lectures delivered by Prof. Parimal Parag in earlier academic years. The LaTeX files of lecture notes were used to create Beamer presentations via Python code. The extracted content from the LaTeX files was used to generate a script for audio generation with the help of LLMs (Gemini). The generated scripts and Beamer presentations were reviewed by a subject expert to ensure correctness. The reviewed scripts were fed to an AI text-tospeech converter (ElevenLabs AI) to generate voiceover for the lectures in Prof. Parimal’s voice. The Beamer presentation slides and AI-generated voiceovers were synced using the ffmpeg package via Python code. In this manner, AI-assisted video lectures were created for all 28 lectures of the Random Process course. We plan to publish this work and make the videos available for free learning purposes.