Konstantinos Kamnitsas MSc PhD

Konstantinos Kamnitsas is Associate Professor of Engineering Science (Biomedical Imaging) at the Department of Engineering Science, and a Non-Tutorial Fellow at Wolfson College. His research focuses on Machine-Learning (ML) and primarily deep neural networks for medical image analysis. His work has two main goals:

• Develop more reliable, transparent and accountable models for safer use in healthcare.
• Empower radiologists, clinicians and researchers with intelligent ML-based tools to better address their research questions and needs of clinical workflows.

Konstantinos completed his PhD in 2019 at Imperial College London, where he developed some of the first 3-dimensional neural networks for processing volumetric medical data, such as MRI and CT, and methods for improving generalization to heterogeneous data. His work has won various awards, among which two international competitions for brain cancer and ischemic stroke lesion segmentation. He also obtained an MSc in Computing Science in 2013, also from Imperial College, and the diploma in Electrical and Computer Engineering in 2010 from Aristotle University of Thessaloniki, Greece. He has also spent time conducting research in industry, such as at the Healthcare Intelligence team of Microsoft Research and Kheiron Medical Technologies.

He became a Lecturer in 2021 at the School of Computer Science of the University of Birmingham, where he retains a position as an Honorary Research Fellow since 2022, when he joined the University of Oxford as an Associate Professor. He is member of the Editorial Board of Medical Image Analysis (MedIA) journal.

View a list of Professor Kamnitsas’ publications on Google Scholar.

We develop machine learning (ML) methods for medical image interpretation and analysis, driven by two main aims:

• To develop more reliable and transparent machine learning models to catalyse safer integration of the technology in real-world applications.
• To facilitate clinical research in a variety of applications (segmentation, detection, reconstruction, etc) and imaging modalities (MRI, CT, Mammography, X-rays, etc).

Therefore we investigate a variety of methodologies such as:

• State-of-the-art neural networks for image understanding and analysis
• Estimating model uncertainty or detecting potential failure of prediction for safe ML (due to corrupted input, unknown pathology, etc)
• Identifying and alleviating bias in a model for fair ML in healthcare (domain adaptation, causality, etc)
• Learning from decentralised data to enable international collaborations (federated learning, etc)
• How to learn useful information from unlabelled data, multi-modal data, and more.

Deep Learning Methods for Medical Image Analysis

We develop neural-network based tools for detection/segmentation of pathologies (tumors, injuries, etc) and tissues of interest for a variety of imaging modalities (MR/CT/Mammography/Xrays/etc). Our aim is to empower clinical researchers with tools to tackle variety of tasks. Among other tools, we have developed and maintain DeepMedic, an open-source, easy-to-use Deep Learning segmentation tool for clinical research.

Detecting Failures of Neural Networks after Deployment

Performance of machine learning models may degrade when they process data that differ from those used during training (distribution shift). This poses a challenge for safe deployment of ML models in medical imaging due to data heterogeneity. We develop methods for detecting distribution shift, uncertainty estimation, and out-of-distribution detection to ensure reliable deployment of ML/AI models. Supported by EPSRC.

Federated Learning

We develop algorithms and optimization techniques that enable training neural networks on decentralised medical databases, held physically at multiple different institutions. Our aim is to enable international collaborations towards learning models that generalize better while preserving privacy of medical data. Supported by EPSRC.

CENTER-TBI

An international collaboration that aims to improve the care for patients with Traumatic Brain Injury (TBI). Our methods contribute in analysing this pathology in MRI and CT data and extract novel insights on the disease.

Are you interested in studying for a PhD (DPhil) in Engineering Science with me? Below are the main routes:

• Direct application for a DPhil at the Department of Engineering Science (via this link). You will need to list me as supervisor. This is the preferred option for students that want to work specifically with me and have defined a research project/direction that fits my interests, which will need to be discussed within their submitted personal statement. PhD duration is 3-4 years, starting next Oct. Note that although all successful applications are considered for scholarships by the Uni/Dept (see prev link), these are limited and very competitive. Applicants are strongly recommended to explore what other external scholarship opportunities they can pursue for supporting their studies.
• Apply for a position at one of the Centers of Doctoral Training (CDT) that I am affiliated with: CDT on Health Data Science (HDS) or CDT on Autonomous Intelligent Machines and Systems (AIMS). This is a 4 year program, where during the 1st year, along with taught courses, students get to choose between projects proposed by affiliated faculty to pursue during the next 3 years for their PhD. CDT websites discuss deadlines and funding of successful applicants.
• Other opportunities will be advertised here when available.

Please have a look at my lab's research areas and my Google Scholar to find my publications and identify my research interests. You can email me for specific follow up questions (e.g. whether a specific research question is of interest if you are considering a direct DPhil application). Please insert “[DPhil EngSci KK]” in your email subject to show you have read the above instructions. I apologise in advance for delays in replying, as the volume of emails can be quite large.