LiFE-Net: Longitudinal information Fusion for Enhanced lesion detection in unsupervised learning contexts

Walid Yassine, Martin Charachon, CELINE HUDELOT, Roberto Ardon

Keywords: Longitudinal imaging, Liver lesion detection, Deep learning, CT scans

Abstract:

Accurate detection of liver lesions in longitudinal follow-up is critical for assessing disease progression. Unlike clinical practices that compare multiple time points, most deep-learning approaches treat these time points independently. Existing longitudinal imaging methods, particularly in brain imaging, use strategies like channel-wise concatenation, recurrent architectures, or temporal difference computation. However, these methods might fall short in liver imaging due to challenges like non-rigid motions, anatomical variability, and changes in imaging conditions.

To address these challenges, we introduce LiFE-Net, the first framework to integrate longitudinal information from baseline liver CT scans through feature fusion. Our method employs intermediate feature fusion via self-attention mechanisms, leveraging baseline images to incorporate longitudinal information for more accurate predictions. We adopt an unsupervised training approach using synthetic lesions to address the lack of supervised datasets for longitudinal liver tumors.

Our results show improvements in detection performance on follow-up images when baseline information is incorporated, with gains in both detection mAP and ROC AUC per exam metrics.
An exhaustive ablation study further highlights the impact of baseline image integration, registration quality, and architectural components in achieving these improvements. Our code for LiFE-Net is made publicly available at: https://github.com/walid-yassine/LiFE-Net

Primary Subject Area: Detection and Diagnosis

Secondary Subject Area: Application: Radiology

Paper Type: Methodological Development

Registration Requirement: Yes

Reproducibility: https://github.com/walid-yassine/LiFE-Net

Visa & Travel: Yes

Midl Latex Submission Checklist: Ensure no LaTeX errors during compilation., Created a single midl25_NNN.zip file with midl25_NNN.tex, midl25_NNN.bib, all necessary figures and files., Includes \documentclass{midl}, \jmlryear{2025}, \jmlrworkshop, \jmlrvolume, \editors, and correct \bibliography command., Did not override options of the hyperref package, Did not use the times package., All authors and co-authors are correctly listed with proper spelling and avoid Unicode characters., Author and institution details are de-anonymized where needed. All author names, affiliations, and paper title are correctly spelled and capitalized in the biography section., References must use the .bib file. Did not override the bibliographystyle defined in midl.cls. Did not use \begin{thebibliography} directly to insert references., Tables and figures do not overflow margins; avoid using \scalebox; used \resizebox when needed., Included all necessary figures and removed *unused* files in the zip archive., Removed special formatting, visual annotations, and highlights used during rebuttal., All special characters in the paper and .bib file use LaTeX commands (e.g., \'e for é)., Appendices and supplementary material are included in the same PDF after references., Main paper does not exceed 9 pages; acknowledgements, references, and appendix start on page 10 or later.

Submission Number: 166