Research Projects

Recent success in many deep learning models has demonstrated significant improvements on various tasks by pretraining on a large repository, followed by finetuning on a particular task. This approach requires a task-specific fine-tuning phase leveraging a good number of examples. On the contrary, humans perform similar process from only a very few samples with better generalizability. For instance, a child can distinguish cats vs dogs even in a few times of observation, whereas it Is a hard task for a machine learning model to learn such concepts with few training samples. Motivated by this rationale, in this proposal, a novel generalized AI model that can depict human-like task-agnostic, few-shot performance with better domain adaptation characteristics is proposed via Meta-learning (or learning to learn). Meta-learning based deep learning strategy is employed towards generalizing well on a given limited amount of training data. This enables the model to quickly learn from fewer samples and to utilize prior knowledge and experience for solving new problems. In the application domains with utmost practical relevance as in the Medical field, such generalizable and multi-task adaptable systems are the need of the hour. In this proposal, a novel meta-learning framework for prediction analysis tasks in magnetic resonance imaging (MRI) for stroke lesions is envisaged. Specifically, a predictive analysis of the final infarct stroke lesions using magnetic resonance images (MRIs) acquired at the initial presentation (baseline) is envisaged by leveraging the meta-learning framework. Some of the State-of-the-art meta-learning strategies and common methods for dynamics forecasting such as Reptile/MAML and U-net respectively will be employed towards this goal. The proposed model can adapt to various settings quickly and enhance various application tasks in medical imaging, without the requirement of a large amount of labelled training data. We contemplate that the proposal will serve as a stepping stone toward the development of such an indigenous meta-learning-based MRI stroke lesion outcome prediction model for the first time in India.