Code for generating synthesis data

[yingqihao2022]

Hi @mrokuss,

Really excited by your amazing idea - generating synthetic longitundial data for augmentation.

Noted that you said in the previous issue (by another guy) that the code is not suitable for public release due to its mess and only suitable for specific lesions. However, I do hope to generate synthetic longitudinal data so as to compare with other synthetic SOTA ideas, such as LeFusion.

Could you just provide a primary example code that is suitable for one specific lesion in any modality, please? It could be only a minority of the whole codebase but including the core steps. Appreciate in advance!

[mrokuss]

Hi @yingqihao2022

Thanks a lot for your kind words and your interest in this idea. I'm glad to hear you're exploring synthetic longitudinal augmentation and comparisons to methods like LeFusion!

Indeed The code was not initially released due to its highly experimental nature, the specificity to the datasets we used, and the heavy reliance on visual inspection for validation. The quality is highly dataset-dependent and not robust across modalities or lesion types without significant manual tuning.

However, based on the effect of the synthetic data in the paper, the positive feedback and to support further experimentation in the community, I've prepared an experimental version of the core code. It's a stripped-down and unrefined version of what was used in the paper, intended to serve as a starting point for exploration rather than a ready-to-use pipeline.

You can download it here !

Please note:

• It's built for nnUNet-style datasets.
• It applies spatial and intensity-based transformations to simulate lesion evolution.
• There's minimal cleanup and no guarantees it will out-of-the-box and beyond the synthetic datasets used in our study.
• Visual inspection remains essential for validating results.

Feel free to reach out if you run into issues or have suggestions. Looking forward to hearing how it works in your setting!

Hope this helps!

Max