Explainable Domain Adaptation in medical Imaging for biomarker extraction using Deep learning

Faculty: Dr. Vaanathi Sundaresan (Computational and Data Science), Dr. Jaya Prakash (Instrumentation and Applied Physics)

Artificial Intelligence (AI) methods applied to medical imaging can lead to imaging biomarkers for differential diagnosis of diseases for personalized treatment. However, the main challenges for AI methods include (1) lack of generalizability across datasets due to heterogeneity in their characteristics across domains (e.g., different scanners/hospitals), and (2) lack of explainability in the decisions made by the model. To address the lack of generalisability, adaptation methods, in general, aim to transfer the knowledge from a source domain to a target domain by leveraging the invariant features across different domains. Also, in clinical settings, data from multiple centres are protected by stringent data sharing/privacy agreements. So, it is crucial that methods are generalised across domains, whilst ensuring data privacy throughout. Similarly, there is a need for robust explainable AI methods that can provide insights regarding their features leading to the biomarker segmentation results, without compromising on their predictive performance across multiple datasets. Further understanding of the working of deep learning models will be highly useful for developing fair, accountable and trustworthy models for biomarker detection in line with regulations (e.g., general data protection regulation) that mandates transparent decision making. 

The aim of this project is to adapt the models for biomarker segmentation on medical images across multiple centres and overcome the main challenges in clinical scenarios: scanner-related shifts and inconsistent set of modalities across studies. The project also aims to develop explainable AI methods to obtain a combination of feature relevance and output probabilities for detection of medical imaging biomarkers for various diseases. Various lesion- /image-level performance metrics will be used to compare the performance of the proposed AI method with existing state-of-the-art methods, and to evaluate their domain-agnostic predictive performance. 

Background needed: Linear Algebra, signal processing, machine learning, programming.

Basic Qualifications: B.E./B.Tech. in EE/ECE/IN/CS/IT/BME (or) M.Sc. (Mathematics/Physics)