Source-linked AI summary
No New-Net
Fabian Isensee, Philipp Kickingereder, Wolfgang Wick, Martin Bendszus, Klaus H. Maier-Hein
TL;DR
Automatic brain-tumor segmentation is challenging because tumors are irregular, while numerous architectural variants complicate identifying broadly useful improvements. This paper instead evaluates a minimally modified U-Net with a training-focused framework and achieves second place in BraTS2018.
Problem
Irregular tumors make automatic brain-tumor segmentation challenging, and the usefulness of numerous architectural modifications across datasets is difficult to assess.
Method
The paper uses a minimally modified 3D U-Net trained with large randomly sampled patches, Dice-based optimization, additional data, region-based training, postprocessing, and combined loss functions.
Results
The approach achieved second place among 64 BraTS2018 teams, with test Dice scores of 77.88, 87.81, and 80.62 for enhancing tumor, whole tumor, and tumor core, respectively.
Takeaways & Limitations
A generic U-Net can obtain very competitive segmentation when its training framework is carefully designed, without significant architectural alterations.
Takeaways & Limitations
The main analysis relies on cross-validation because the validation set contains 66 cases versus 285 training cases, and training-set conclusions may be overestimated when using additional data.
Abstract
from arXiv · showhide
In this paper we demonstrate the effectiveness of a well trained U-Net in the context of the BraTS 2018 challenge. This endeavour is particularly interesting given that researchers are currently besting each other with architectural modifications that are intended to improve the segmentation performance. We instead focus on the training process arguing that a well trained U-Net is hard to beat. Our baseline U-Net, which has only minor modifications and is trained with a large patch size and a Dice loss function indeed achieved competitive Dice scores on the BraTS2018 validation data. By incorporating additional measures such as region based training, additional training data, a simple postprocessing technique and a combination of loss functions, we obtain Dice scores of 77.88, 87.81 and 80.62, and Hausdorff Distances (95th percentile) of 2.90, 6.03 and 5.08 for the enhancing tumor, whole tumor and tumor core, respectively on the test data. This setup achieved rank two in BraTS2018, with more than 60 teams participating in the challenge.
1 Introduction
The paper addresses challenging automatic brain-tumor segmentation by testing whether a well-trained, minimally modified U-Net can remain competitive amid extensive architectural innovation.
- Automatic segmentation could make tumor assessment faster, more objective, and potentially more accurate, but irregular tumor structure makes the task challenging.
- BraTS 2018 provides annotated low- and high-grade glioma cases from diverse institutions, scanners, field strengths, and acquisition protocols.
- BraTS evaluates participating algorithms using Dice coefficient, sensitivity, specificity, and 95th-percentile Hausdorff distance.
- CNNs dominate biomedical image segmentation and recent BraTS winners, with architectures such as DeepMedic combining low- and high-resolution pathways.
- The authors report that architectural variants provided no additional benefit over a well-trained U-Net and use this minimally altered model to pursue competitive BraTS 2018 segmentation.
2 Methods
The method uses a minimally modified 3D U-Net and emphasizes preprocessing, training optimization, augmentation, region-based objectives, additional data, and targeted postprocessing.
- Network architecture: The approach instantiates a 3D U-Net close to the original while optimizing its training procedure for BraTS 2018 data.
- Network architecture: The network processes 128x128x128 patches with batch size 2, reduces filters before upsampling, and reaches 30 feature channels at the highest resolution.
- Network architecture: Leaky ReLUs and instance normalization replace standard ReLUs and batch normalization to improve behavior with the chosen loss and small batch size.
- Training procedure: Training uses random patches, ADAM, learning-rate reduction, early stopping, and l2 weight decay of 10^-5.
- Training procedure: A multiclass soft Dice loss addresses class imbalance, using softmax outputs and one-hot ground-truth labels across patch pixels and classes.
- Training procedure: The pipeline applies on-the-fly augmentation, whole-patient inference with mirrored test-time augmentation, region-based training, and cotraining on public or institutional data.
- Cotraining: Cotraining uses separate dataset-specific segmentation layers while sharing the rest of the network, accommodating potentially different label definitions.
- Loss and postprocessing: The method targets false-positive enhancing-tumor predictions and combines Dice with cross-entropy to address rare-class and probability-calibration issues.
3 Experiments and Results
The experiments evaluate a carefully trained U-Net using cross-validation, validation ensembles, qualitative inspection, and final test-set comparison. Additional data, postprocessing, and training choices produced strong validation and second-place test performance.
- Experimental setup: Five-fold cross-validation was run on the 285 BraTS 2018 training cases, while validation results used an ensemble of the five cross-validation networks.The authors emphasize training-set cross-validation because the validation set contains only 66 cases.
- Training results: 73.43/89.76/82.17 Dice scores (enh/whole/core) were achieved by the baseline U-Net on the training set.The baseline used large input patches and a Dice loss function.
- Validation results: 80.87/91.26/86.34 Dice scores (enh/whole/core) and 2.41/4.27/6.52 Hausdorff distances were obtained on the validation set by an ensemble of models using additional data.The ensemble combined models cotrained with decathlon and institutional data.
- Test results: The final submission used an ensemble of two five-model ensembles, producing 10 predictions per test case and achieving second place among 64 teams.The method had strong enhancing-tumor results, while the winning contribution outperformed it for tumor core and whole tumor.
- Qualitative results: The qualitative example shows FLAIR, T1ce, and segmentation views, with enhancing tumor, necrosis, and edema encoded in yellow, turquoise, and violet.The case is patient CBICA AZA 1 from the validation set.
4 Discussion
The paper concludes that a generic U-Net with minor modifications can achieve competitive segmentation when its training procedure is carefully designed. Region-based training, additional data, postprocessing, and combined losses substantially improved the model, which placed second in BraTS2018.
- Discussion: A generic U-Net with only minor modifications achieved very competitive segmentation when trained correctly.The conclusion attributes the result to the training framework rather than a highly specialized architecture.
- Discussion: Region-based training, additional-data cotraining, targeted postprocessing, and combined Dice and cross-entropy losses substantially increased performance.These measures enhanced the training procedure beyond the already strong base model.
- Discussion: The final ensemble combined models cotrained with public and institutional data and achieved second place in the BraTS2018 challenge.The result is presented as evidence of the impact of a well-designed training framework.