Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Uncertain but Useful: Leveraging CNN Variability into Data Augmentation

Created by
  • Haebom

Author

In es Gonzalez-Pepe, Vinuyan Sivakolunthu, Yohan Chatelain, Tristan Glatard

Outline

This paper investigates numerical instability arising during the training of FastSurfer, a deep learning (DL)-based brain imaging analysis pipeline. We analyze the variability of FastSurfer's training process using controlled perturbations using floating-point perturbations and random seeds, demonstrating that DL is more susceptible to instability than conventional neuroimaging pipelines. However, the ensemble generated through perturbations performs similarly to the baseline model without perturbations, demonstrating that this variability can be leveraged for subsequent applications such as brain age regression analysis. Our conclusion suggests that training-time variability is not only a reproducibility issue but can also be leveraged as a resource to enhance robustness and enable new applications.

Takeaways, Limitations

•
Takeaways:
◦
We identified numerical instabilities that occur during the training process of a deep learning-based brain imaging analysis pipeline and analyzed their causes and impacts.
◦
We exploit the variability in the training process to create ensemble models, suggesting the potential for performance enhancement and the development of new applications.
◦
We present a case study demonstrating that variability in the training process can be leveraged as a data augmentation strategy.
•
Limitations:
◦
As this analysis is for a single pipeline, FastSurfer, generalizability to other DL-based brain imaging analysis pipelines is limited.
◦
Brain age regression analysis Results for only one application are presented, so applicability to other applications requires further study.
◦
The types and strengths of perturbations used in the analysis are limited, so further research on other types of perturbations is needed.
👍
Made with Slashpage