Learning deep abdominal CT registration through adaptive loss weighting and synthetic data generation

TL;DR

This paper introduces a training framework for deep abdominal CT registration that combines synthetic data generation, adaptive loss weighting, and transfer learning, leading to improved registration accuracy without increasing inference time.

Contribution

It proposes a novel training strategy incorporating synthetic data, dynamic loss weighting, and transfer learning to enhance deep image registration performance.

Findings

Segment-guided registration improves accuracy.

Transfer learning from brain MRI to abdominal CT is effective.

Dynamic loss weighting marginally boosts performance.

Abstract

Purpose: This study aims to explore training strategies to improve convolutional neural network-based image-to-image deformable registration for abdominal imaging. Methods: Different training strategies, loss functions, and transfer learning schemes were considered. Furthermore, an augmentation layer which generates artificial training image pairs on-the-fly was proposed, in addition to a loss layer that enables dynamic loss weighting. Results: Guiding registration using segmentations in the training step proved beneficial for deep-learning-based image registration. Finetuning the pretrained model from the brain MRI dataset to the abdominal CT dataset further improved performance on the latter application, removing the need for a large dataset to yield satisfactory performance. Dynamic loss weighting also marginally improved performance, all without impacting inference runtime. Conclusion: Using simple concepts, we improved the performance of a commonly used deep image registration architecture, VoxelMorph. In future work, our framework, DDMR, should be validated on different datasets to further assess its value.