MAMA-MIA-Challenge (Team FME)

This repository collects the preprocessing, training and inference code for the MAMA-MIA challenge submission of team FME.

Team Members: Kai Geissler, Jan Raphael Schäfer

Final Submission

Inference code

The inference code of our final submission for both tasks is in the directory inference/code_submission_7.

Training code

The data preparation for the segmentation task is the script data_preprocessing_segmentation/copy_and_compute_first_subtraction_images.py which we used to compute the first subtraction images from the MAMA-MIA dataset, which we than trained a 5-fold ensemble of nnU-Nets on:

nnUNetv2_plan_experiment -d 125 -pl nnUNetPlannerResEncL -c 3d_fullres
nnUNetv2_train 125 3d_fullres 0,1,2,3,4 -p nnUNetResEncUNetLPlans --c

The data preparation for the classification task was done using the script data_prepocessing_classification/crop_and_copy_first_three_timepoints.py. The training was performed with the script hyperparameter_tuning/tune_random_search_3d.py.

Description of segmentation approach

The segmentation approach employed a 5-fold ensemble of nnU-Net models configured with the 3d_fullres setup. Data preprocessing involved computing the first difference DCE-MRI image, while nnU-Net handled normalization, resampling, and patch generation automatically. During inference, test-time augmentation via mirroring was applied, and connected component analysis was used as post-processing to retain only the largest connected lesion. The nnU-Net configuration included a voxel size of 2.0x0.7x0.7, a patch size of 80x256x256, and a batch size of 2.

Description of classification approach

The classification approach used 3D models pretrained for video classification from the TorchVision library. The first three DCE-MRI images were selected as input, z-score normalized using the pre-contrast image, cropped to a bounding box around the lesion, and resampled to a fixed size. A hyperparameter search with Ray Tune optimized parameters such as learning rate, weight decay, and model architecture, resulting in an ensemble of 25 models trained with 5-fold cross-validation. Test-time augmentation was applied by flipping the input images along each axis, producing 8 augmented versions of each image, with predictions averaged across them. The classification models were trained using the AdamW optimizer, cosine annealing learning rate scheduler, and CCE loss function.

Set up training dependencies

Install your favorite python package manager (e.g. miniconda).

conda create -n 'mamamia' python==3.12.9
conda activate mamamia
pip install torch==2.7.1 torchvision==0.22.1 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt

Test inference Docker container

docker run -it --rm --entrypoint /bin/bash --gpus all --shm-size=1024m `
--mount type=bind,source=<this_repository_checkout_directory>/mama-mia-challenge/inference,target=/app/ingested_program `
--mount type=bind,source=<test_data_directoy>,target=/dataset `
--mount type=bind,source=<test_output_directory>,target=/output `
lgarrucho/codabench-gpu:latest `
 -c "python /app/ingested_program/test_model.py /dataset /output"

Folder structure

• breast_fgt_segmentation: Segment breast and FGT in breast MRI
• classification_training: Train classification models
• data_prepocessing_classification: Preprocessing of classification data
• data_preprocessing_segmentation: Preprocessing of segmentation data
• hyperparameter_tuning: Hyperparameter tuning for classification models
• inference: Submission code
• tests: Small tests for selected components