Team

Ihab Asaad, Nathalie Demme, Tim Büchner

Overview

Incomplete tumor resection in the head and neck region remains a major challenge in oncologic surgery because the close proximity to vital structures often precludes the generous safety margins standard in other body parts. This creates a high risk of leaving microscopic residual tumor tissue behind, which is known to significantly decrease patient survival rates and increase the risk of recurrence. The Sensorized Surgery project aims to develop a sensor-equipped surgical system that integrates multimodal marker-free imaging, mechanical tissue sensing, and AI-driven analysis. The goal is to continuously predict tumor boundaries and provide visual and haptic feedback, enabling more precise surgery.

Our task in the Computer Vision Group is to develop high-performance semantic segmentation models that accurately distinguish tumors from surrounding tissues, maintaining robustness against distribution shifts arising from varying clinical conditions.

Stable Features

To ensure robust and reliable performance during clinical deployment, one of our work packages focuses on addressing distribution shifts that arise from discrepancies between offline training data and real surgical environments. Such shifts are often caused by dataset biases, including spurious correlations and class imbalance. Our work therefore centers on developing debiasing strategies that promote the learning of stable and invariant features, reducing the model’s reliance on confounding factors and improving generalization across diverse clinical conditions.

Robust Models

The focus lies on optimizing neural network architectures and implementing advanced training strategies to significantly improve segmentation accuracy. A central challenge involves addressing sample scarcity, as high-quality, annotated medical data for head and neck tumors is often limited. To ensure the models remain robust under these conditions, strategies are developed to achieve high performance even with small datasets. Furthermore, the research tackles the problem of noisy labels, which arise from the inherent difficulty of precisely marking tumor boundaries in complex surgical images. By refining models to account for these labeling uncertainties and data constraints, the goal is to achieve more precise tissue differentiation, providing the surgeon with a reliable tool to maximize resection success and patient survival.

Links

Project Webpage