https://atlas.rsna.org/schemas/2025-11/model.json

3D MRI brain tumor segmentation using autoencoder regularization

https://arxiv.org/abs/1810.11654

BraTS 2018 winner

amyronenko@nvidia.com

This approach won 1st place in the BraTS 2018 challenge for brain tumor segmentation.

Semantic segmentation network based on encoder-decoder architecture with an asymmetrically large encoder and a variational auto-encoder (VAE) branch for regularization. Uses ResNet blocks with Group Normalization.

The model won the BraTS 2018 challenge. Implementation in Tensorflow. Trained on NVIDIA Tesla V100 32GB GPU.

Automated segmentation of 3D brain tumors can save physicians time and provide an accurate reproducible solution for further tumor analysis and monitoring.

Diagnosis, monitoring, and treatment planning

Fully automated

Segmentation of brain tumor subregions (whole tumor, tumor core, enhancing tumor) from multimodal 3D MRI scans for patients with primary or secondary brain tumors, including gliomas (low-grade and high-grade).

Four channel 3D MRI modalities (T1, T1c, T2, FLAIR) rigidly aligned, resampled to 1x1x1 mm isotropic resolution, skull-stripped, and normalized to zero mean and unit std. Input image size 240x240x155, processed with random crop of 160x192x128.

Input images must be 3D multimodal MRI (T1, T1c, T2, FLAIR), rigidly aligned, resampled to 1x1x1 mm isotropic resolution, skull-stripped, and normalized. Test time augmentation by mirror flipping input 3D image axes and averaging 8 flipped segmentation probability maps can be applied. Ensemble of 10 models can be used for improved performance.

Performance may be affected by data outside the distribution of the BraTS 2018 dataset. Batch size of 1 was used due to GPU memory limits, which required a smaller crop size compared to some related works. More sophisticated data augmentation techniques (histogram matching, affine transforms, filtering) and post-processing (CRF) did not consistently improve performance.

The VAE branch helped to consistently achieve good training accuracy for any random initialization.

Training a single model for 300 epochs takes 2 days on a single NVIDIA Tesla V100 32GB GPU. Inference time is 0.4 sec for a single model on a single V100 GPU.