Master's thesis - Monitoring and Mitigating AI Model Degradation in Surgical Workflow

14.08.2025, Abschlussarbeiten, Bachelor- und Masterarbeiten

The aim of this thesis is to investigate model degradation in AI systems for laparoscopic surgery and develop strategies to detect and compensate for shifts in surgical data distributions.

Background:
Artificial intelligence (AI) has shown great potential in enhancing surgeries by decision support. However, models trained on historical surgical data often face performance degradation over time due to shifts in surgical workflows, instruments, or patient populations. Such model drift or data distribution shifts can reduce reliability and safety in clinical settings. Understanding, detecting, and mitigating these shifts is therefore critical to maintain robust AI performance in the operating room.

Task:
The aim of this thesis is to investigate model degradation in AI systems for laparoscopic surgery and develop strategies to detect and compensate for shifts in surgical data distributions. Using a rich dataset of historical laparoscopic surgeries collected over 10 years, the thesis will focus on workflow analysis and/or instrument tracking and segmentation, exploring how AI models evolve and where performance might decline over time. The work will involve both quantitative analysis of model performance and qualitative insights from clinical workflows to propose robust, adaptive methods for AI in surgery.

Subtasks:
Within the scope of the thesis, the following points should be addressed:
• Literature review on model drift, dataset shift, and robustness in AI for medical imaging and surgical applications
• Analysis of historical laparoscopic surgery data to characterize potential sources of model degradation
• Evaluation of AI models for instrument tracking, segmentation, or workflow analysis over time
• Development of strategies to detect, quantify, and mitigate model degradation (e.g., retraining protocols, continual learning, or domain adaptation)
• Validation of proposed strategies on retrospective data and, if possible, in collaboration with clinical partners

Your Profile:
• Studies in a field relevant to the task description (e.g., computer science, medical engineering, data science)
• Strong background in machine learning, computer vision, or AI
• Proficiency in Python and deep learning frameworks such as PyTorch
• Interest in medicine and understanding of surgical workflows

Application:
Please send your application with the listed attachments to one of the following addresses: lars.wagner@tum.de, dennis.schneider@tum.de
Attachments: CV, Transcript of Records, Brief statement of motivation, including relevant background

Kontakt: lars.wagner@tum.de

Mehr Information

https://www.pm.mh.tum.de/miti/ausschreibungen/monitoring-and-mitigating-ai-model-degradation-in-surgical-workflow/