Evaluation of Unsupervised Deep Learning-based Methods for Chest and Abdominal CT Image Registration

Laura Estacio-Cerquin, Daniël Boeke, Kalina Chupetlovska, Regina Beets-Tan, Stefan Klein, Stefano Trebeschi

Keywords: Image Registration, Deep Learning, Chest CT, Abdominal CT, Evaluation

TL;DR: Systematic comparison of common deep learning methods for image registration in CT scans

Abstract: Image registration is crucial for the alignment of medical images, enabling better analysis and interpretation. Although deep learning-based methods have shown promising results, the impact of architectural choices remains unclear, especially when training on scarce, small, or low-quality datasets. This study compares four different architectures for deep learning-based image registration. All methods were trained in an unsupervised setting, using the same loss function and optimization method, but with optimized hyperparameters for each method. In addition, two conventional optimization-based methods were included in the comparison. Experiments were performed on Lung and Abdomen CT datasets of the Learn2Reg challenge. Our findings suggest that the performance of deep learning-based methods varies substantially depending on the dataset type and its specific challenges.

Submission Number: 3