LesiOnTime: Joint Temporal and Clinical Modeling for Small Breast Lesion Segmentation in Longitudinal DCE-MRI

How LesiOnTime Works?

Installation

LesiOnTime builds on two state-of-the-art baselines—nnU-Net for non-longitudinal segmentation and LongiSeg for longitudinal segmentation. To get started, please follow the installation instructions in the LongiSeg repository.

On top of these frameworks, LesiOnTime adds two key innovations at every network layer:

1. Temporal Prior Attention (TPA)
   A specialized attention module that leverages temporal changes across successive DCE-MRI scans to highlight evolving lesion characteristics.

2. BI-RADS Consistency Regularizer (BCR) Loss
   A latent-space regularization term that enforces consistency with BI-RADS clinical assessments, improving robustness and interpretability.

Together, these modules deliver more accurate, clinically grounded segmentation of small breast lesions in longitudinal imaging studies.

Usage: How to Train LesiOnTime with TPA and BCR?

1. Update the Codebase

   • Modify LesiOnTime & TPA classes from

     LesiOnTime/architectures/longi_unet_difference_weighting.py
     

   • Move compound_losses.py to

     LongiSeg/longiseg/training/loss/
     

   • Update DifferenceWeightingBlock from

     LesiOnTime/building_blocks/difference_weighting_block.py
     

2. Train LesiOnTime with TPA module only

   • Add the trainer file: trainers/LesiOnTime_TPA_Trainer.py to LongiSeg/longiseg/training
   • Run training:

     LongiSeg_train DATASET_NAME_OR_ID 3d_fullres "all" -tr LesiOnTime_TPA_Trainer
     

3. Train LesiOnTime with BCR loss only

   • Add the trainer file: LesiOnTime_BCR_Trainer.py to LongiSeg/longiseg/training
   • Run training:

     LongiSeg_train DATASET_NAME_OR_ID 3d_fullres "all" -tr LesiOnTime_BCR_Trainer
     

4. Train LesiOnTime with both TPA module & BCR loss

   • Add the trainer file: LesiOnTime_TPA_BCR_Trainer.py to LongiSeg/longiseg/training
   • Run training:

     LongiSeg_train DATASET_NAME_OR_ID 3d_fullres "all" -tr LesiOnTime_TPA_BCR_Trainer
     

Citation

@inproceedings{Kamran2025LesiOnTimeJ,
  title  = {LesiOnTime -- Joint Temporal and Clinical Modeling for Small Breast Lesion Segmentation in Longitudinal DCE-MRI},
  author = {Mohammed Kamran and Maria Bernathova and Raoul Varga and Christian F Singer and Zsuzsanna Bago-Horvath and Thomas Helbich and Georg Langs and Philipp Seebock},
  year   = {2025},
  url    = {https://api.semanticscholar.org/CorpusID:280417197}
}

Acknowledgements

This work was supported by the European Federation for Cancer Images (EUCAIM, Grant No. 101100633), Initiative Krebsforschung – Grant 2022 (UE77104006), and Vienna Science and Technology Fund (WWTF, PREDICTOME) [10.47379/LS20065].