Applying Longitudinal Augmentation and Data Generation (LAUGEN) in Medical Imaging

Nico Disch, Balint Kovacs, Constantin Ulrich, Robin Peretzke, Saikat Roy, Maximilian Rouven Rokuss, Yannick Kirchhoff, David Zimmerer, Klaus Maier-Hein

Keywords: Augmentations, Synthetic Data, Medical Imaging, Longitudinal Data

TL;DR: LAUGEN is a lightweight framework, which generates semi-synthetic image time series from a single image and a segmentation, enabling data augmentation and data generation for model testing in medical longitudinal tasks.

Abstract:

Deep learning had transformative impact in medical imaging, in areas such as classification, segmentation, and report generation.
Another area holding promises, is the area of personalized medicine, especially disease progression modeling.
However, longitudinal imaging is even more data-constrained than single time-point imaging, as it requires repeated acquisitions over extended periods, often spanning months or even years.
To address this challenge, we introduce Longitudinal Augmentation and Data Generation (LAUGEN) a lightweight, semi-synthetic image generation framework which can be applied in the domain of medical image time series.
LAUGEN is efficient, requiring only a single image and its segmentation to produce diverse pseudo-temporal sequences, and is capable of handling typical $3D$ medical data.
We demonstrate its use as a data augmentation strategy for improving model performance and propose its role as a tool for unit testing longitudinal models, where pre-defined latent progressions enable controlled and arbitrarily many evaluations.
Our qualitative results on the Brain Tumor Segmentation (BraTS) dataset and quantitative experiments on Automated Cardiac Diagnosis Challenge (ACDC) dataset highlights \LA{}'s potential to enrich datasets and enhance result diversity.

Submission Number: 39