Interactive Segmentation and Annotation of Medical Images

Hybrid Talk:

ATV SR ground floor A0.106, INF242

Zoom Registration:

https://dkfz-de.zoom-x.de/meeting/register/u5EsceCqqjsoGtYp-yUlOXWqk8Z9U1g6vitH 

Interactive segmentation is an emerging field in medical image analysis, aimed at reducing the time and effort required for manual annotations by incorporating real-time human feedback. By leveraging user inputs—such as clicks, scribbles, or masks—interactive segmentation allows for the iterative refinement of model predictions, effectively guiding the system toward greater accuracy and efficiency. Recent advancements in deep learning have significantly accelerated progress in this domain, with over 120 methods proposed specifically for medical imaging in the past five years.
This talk will delve into how user interactions are transformed into guidance signals that direct models toward more precise outcomes. These interactions are often simulated during training to mirror real-world user behaviour, ensuring that models are robust and adaptable in practical scenarios. Additionally, the role of active learning will be examined, showcasing how models intelligently select the most informative samples, thereby reducing the annotation workload while optimizing model performance.
The discussion will also address critical challenges in the field, such as the lack of standardized benchmarks and inconsistent evaluation practices across methods. By offering a structured overview of current approaches, the talk will highlight emerging opportunities and explore the future of interactive segmentation in medical imaging, particularly in the era of foundation models.

Zdravko Marinov received his bachelor's and master's degree in computer science from the Karlsruhe Institute of Technology, Germany, in 2019 and 2021 respectively. He is currently working toward a PhD in the Computer Vision for Human-Computer Interaction lab at the Karlsruhe Institute of Technology under the Helmholtz Information & Data Science School for Health (HIDSS4Health) and the supervision of Prof. Rainer Stiefelhagen. His research interests include interactive segmentation using deep learning to accelerate the annotation of medical images and medical image analysis. He has a strong collaboration with the Institute of Artificial Intelligence (IKIM) led by Prof. Jens Kleesiek in the University Clinic Essen. In the last two years he has published his research in TPAMI, ICCV, ECCV, IROS, MICCAI, among others, and has landed top spots in segmentation challenges in MICCAI and CVPR.

Contact details: https://cvhci.anthropomatik.kit.edu/people_2240.php