Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Automated Estimation of Anatomical Risk Metrics for Endoscopic Sinus Surgery Using Deep Learning

Konrad Reuter, Lennart Thaysen, Bilkay Doruk, Sarah Latus, Brigitte Holst, Benjamin Becker, Dennis Eggert, Christian Betz, Anna-Sophie Hoffmann, Alexander Schlaefer

TL;DR

The paper tackles the time-consuming preoperative risk assessment for endoscopic sinus surgery by automatically estimating Keros, Gera, and TMS scores from coronal CBCT slices. It introduces a heatmap regression-based pipeline with a global-to-local two-stage strategy to localize ten skull-base landmarks, comparing it against a direct single-stage approach across multiple architectures. The results show that the global-to-local method improves robustness and reduces prediction error, achieving Keros MAEs around $0.5$–$0.6$ mm and Gera/TMS estimates with competitive accuracy, while highlighting class-imbalance effects. This automated approach can streamline preoperative planning and support consistent risk stratification across centers, with future work targeting 3D extension and broader multi-center validation.

Abstract

Endoscopic sinus surgery requires careful preoperative assessment of the skull base anatomy to minimize risks such as cerebrospinal fluid leakage. Anatomical risk scores like the Keros, Gera and Thailand-Malaysia-Singapore score offer a standardized approach but require time-consuming manual measurements on coronal CT or CBCT scans. We propose an automated deep learning pipeline that estimates these risk scores by localizing key anatomical landmarks via heatmap regression. We compare a direct approach to a specialized global-to-local learning strategy and find mean absolute errors on the relevant anatomical measurements of 0.506mm for the Keros, 4.516° for the Gera and 0.802mm / 0.777mm for the TMS classification.

Automated Estimation of Anatomical Risk Metrics for Endoscopic Sinus Surgery Using Deep Learning

TL;DR

The paper tackles the time-consuming preoperative risk assessment for endoscopic sinus surgery by automatically estimating Keros, Gera, and TMS scores from coronal CBCT slices. It introduces a heatmap regression-based pipeline with a global-to-local two-stage strategy to localize ten skull-base landmarks, comparing it against a direct single-stage approach across multiple architectures. The results show that the global-to-local method improves robustness and reduces prediction error, achieving Keros MAEs around – mm and Gera/TMS estimates with competitive accuracy, while highlighting class-imbalance effects. This automated approach can streamline preoperative planning and support consistent risk stratification across centers, with future work targeting 3D extension and broader multi-center validation.

Abstract

Endoscopic sinus surgery requires careful preoperative assessment of the skull base anatomy to minimize risks such as cerebrospinal fluid leakage. Anatomical risk scores like the Keros, Gera and Thailand-Malaysia-Singapore score offer a standardized approach but require time-consuming manual measurements on coronal CT or CBCT scans. We propose an automated deep learning pipeline that estimates these risk scores by localizing key anatomical landmarks via heatmap regression. We compare a direct approach to a specialized global-to-local learning strategy and find mean absolute errors on the relevant anatomical measurements of 0.506mm for the Keros, 4.516° for the Gera and 0.802mm / 0.777mm for the TMS classification.

Paper Structure

This paper contains 6 sections, 1 equation, 2 figures, 5 tables.

Table of Contents

  1. INTRODUCTION
  2. MATERIAL AND METHODS
  3. Dataset
  4. Deep Learning Approach
  5. RESULTS
  6. DISCUSSION AND CONCLUSION

Figures (2)

  • Figure 1: Visual Representation of the Keros (left), Gera (middle) and TMS (right) Score. The red dots represent the points to be predicted by our models.
  • Figure 2: Method overview. A global model (a) is used to generate position estimates of the center point of the Keros and Gera scores (red box), as well as the left and right orbital floor (blue boxes). Subsequently, two local models operate on patches extracted around the global position estimates. The first local model (b) estimates the anatomical landmarks for Keros and Gera. The second local model (c) refines the orbital floor position estimates.