Learning-based Inference of Longitudinal imAge Changes (LILAC)

Overview

LILAC is a model that learns to compare images, designed to automatically filter out nuisance variations and focus on clinically significant changes in longitudinal imaging data.

This repository includes models and train/evaluation code for the LILAC paper.

For more information, please read the PNAS paper (in-print, 2025). For questions and feedback, please open issues or email the corresponding author.

Installation

git clone https://github.com/heejong-kim/lilac.git
cd lilac
# if you prefer to use install in conda env
conda create -n lilac python pip 
conda activate lilac
pip install -e .

# If you would like to obtain the trained model weights described in the paper:
wget https://zenodo.org/records/14713287/files/lilac_model_weights.tar.gz
tar -xvzf ./lilac_model_weights.tar.gz

Requirements

The LILAC package depends on the following requirements:

• torch
• torchvision>=0.11.2
• torchio>=0.18.73
• pandas>=1.3.4
• pillow>=8.4.0
• numpy>=1.21.2
• tensorboard
• scikit-learn

Running pip install -e . will automatically check for and install all of these requirements.

All baseline experiments in the paper were run on Python 3.9.7 and CUDA 11.2.

Examples

Learning to temporally order embryo images (LILAC-o)

python run.py \
    --jobname='embryo' \
    --task_option='o' \
    --targetname='phaseidx' \
    --backbone_name='cnn_2D' \
    --batchsize=64 \
    --max_epoch=40 \
    --output_directory='./output' \
    --image_directory='./embryo_image_location' \
    --image_size='224,224' \
    --csv_file_train='./demo_for_release/demo_embryo_train.csv' \
    --csv_file_val='./demo_for_release/demo_embryo_val.csv' \
    --csv_file_test='./demo_for_release/demo_embryo_test.csv'

Learning to temporally order wound-healing images (LILAC-o)

python run.py \
    --jobname='woundhealing' \
    --task_option='o' \
    --targetname='timepoint' \
    --backbone_name='cnn_2D' \
    --batchsize=128 \
    --max_epoch=40 \
    --output_directory='./output' \
    --image_directory='./woundhealing_image_location' \
    --image_size='224,224' \
    --csv_file_train='./demo_for_release/demo_woundhealing_train.csv' \
    --csv_file_val='./demo_for_release/demo_woundhealing_val.csv' \
    --csv_file_test='./demo_for_release/demo_woundhealing_test.csv'

Learning to predict temporal changes (LILAC-t)

python run.py \
    --jobname='oasis-aging' \
    --task_option='t' \
    --targetname='age' \
    --backbone_name='cnn_3D' \
    --batchsize=16 \
    --max_epoch=100 \
    --output_directory='./output' \
    --image_directory='./oasis-aging_image_location' \
    --image_size='128,128,128' \
    --csv_file_train='./demo_for_release/demo_oasis-aging_train.csv' \
    --csv_file_val='./demo_for_release/demo_oasis-aging_val.csv' \
    --csv_file_test='./demo_for_release/demo_oasis-aging_test.csv'

Learning to predict specific changes (LILAC-s)

python run.py \
    --jobname='adni-mci' \
    --task_option='s' \
    --targetname='CDRSB' \
    --optional_meta='AGExSEX'\
    --backbone_name='cnn_3D' \
    --batchsize=16 \
    --max_epoch=40 \
    --output_directory='./output' \
    --image_directory='./adni-mci_image_location' \
    --image_size='128,128,128' \
    --csv_file_train='./demo_for_release/demo_adni-mci_train.csv' \
    --csv_file_val='./demo_for_release/demo_adni-mci_val.csv' \
    --csv_file_test='./demo_for_release/demo_adni-mci_test.csv'

Inference

Simply adding --run_mode='eval' will run inference on the testing set specified in the CSV file provided through --csv_file_test=testset_list.csv. For example,

python run.py \
    --jobname='adni-mci' \
    --task_option='s' \
    --targetname='CDRSB' \
    --optional_meta='AGExSEX' \
    --backbone_name='cnn_3D' \
    --batchsize=16 \
    --max_epoch=40 \
    --output_directory='./output' \
    --image_directory='./adni-mci_image_location' \
    --image_size='128,128,128' \
    --csv_file_train='./demo_for_release/demo_adni-mci_train.csv' \
    --csv_file_val='./demo_for_release/demo_adni-mci_val.csv' \
    --csv_file_test='./demo_for_release/demo_adni-mci_test.csv' \
    --run_mode='eval' \
    # --pretrained_weight ## uncomment this flag if you want to use pretrained weights
  

Refer to ./script-to-replicate-paper-results.sh to run the evaluation on pre-trained weights. All *.csv files above are available in ./demo_for_release.