Source-linked AI summary
Brain Tumor Segmentation and Radiomics Survival Prediction: Contribution to the BRATS 2017 Challenge
Fabian Isensee, Philipp Kickingereder, Wolfgang Wick, Martin Bendszus, Klaus H. Maier-Hein
TL;DR
Manual brain-tumor segmentation is laborious and subjective, while automatic segmentation remains difficult because tumors are irregular. The paper develops a U-Net-inspired convolutional network with Dice loss and augmentation, and combines segmentation-derived radiomics with ensemble regression for survival prediction. It reports strong BraTS segmentation results and a survival-prediction contribution, including 0.858 mean Dice for whole tumor on the test set.
Problem
Automatic brain-tumor segmentation is challenging despite the clinical value of quantitative tumor analysis, and survival prediction is difficult with limited patient information.
Method
The paper develops a U-Net-inspired convolutional segmentation network trained with Dice loss and augmentation, alongside radiomics features and an ensemble regressor for survival prediction.
Results
The model achieved mean Dice scores of 0.858 for whole tumor, 0.775 for tumor core, and 0.647 for enhancing tumor on the BraTS 2017 test set.
Takeaways & Limitations
The contribution combines a competitive BraTS segmentation system with a radiomics-based survival-prediction approach using image-derived tumor features.
Takeaways & Limitations
The tumor label integrated into necrosis for BraTS 2017 was often poorly defined in the training data, leading the algorithm to infer its placement from context.
Abstract
from arXiv · showhide
Quantitative analysis of brain tumors is critical for clinical decision making. While manual segmentation is tedious, time consuming and subjective, this task is at the same time very challenging to solve for automatic segmentation methods. In this paper we present our most recent effort on developing a robust segmentation algorithm in the form of a convolutional neural network. Our network architecture was inspired by the popular U-Net and has been carefully modified to maximize brain tumor segmentation performance. We use a dice loss function to cope with class imbalances and use extensive data augmentation to successfully prevent overfitting. Our method beats the current state of the art on BraTS 2015, is one of the leading methods on the BraTS 2017 validation set (dice scores of 0.896, 0.797 and 0.732 for whole tumor, tumor core and enhancing tumor, respectively) and achieves very good Dice scores on the test set (0.858 for whole, 0.775 for core and 0.647 for enhancing tumor). We furthermore take part in the survival prediction subchallenge by training an ensemble of a random forest regressor and multilayer perceptrons on shape features describing the tumor subregions. Our approach achieves 52.6% accuracy, a Spearman correlation coefficient of 0.496 and a mean square error of 209607 on the test set.
1 Introduction
The paper addresses the challenge of automatic brain-tumor segmentation for clinically useful quantitative analysis and presents a U-Net-based BraTS 2017 contribution combining architectural modifications, Dice loss, and data augmentation. It also applies radiomics with an ensemble model to survival prediction.
- 1 Introduction: Automatic segmentation could make tumor quantification faster, more objective, and potentially more accurate, but irregular tumor structure makes the task challenging.Relevant parameters include tumor-subregion volumes.
- 1 Introduction: The BraTS 2017 dataset contains 210 HGG and 75 LGG cases with multimodal MRI and expert-approved manual segmentations.The supplied passage lists T1-weighted, post-contrast T1-weighted, T2-weighted, and FLAIR MRI.
- 1 Introduction: The proposed segmentation network is based on U-Net but changes its pathway design, normalization, feature-map counts, nonlinearities, and upsampling structure.The authors particularly emphasize optimizing feature maps in the localization pathway.
- 1 Introduction: The method uses a multiclass Dice loss and extensive data augmentation to address class imbalance and reduce overfitting.The contribution passage identifies both choices as central design elements.
- 1 Introduction: For survival prediction, the authors use radiomic tumor features with a random forest regressor and multilayer perceptron ensemble.The features are derived from segmentation-based tumor characterization.
2 Methods
The methods combine multimodal MRI preprocessing, a 3D U-Net-inspired segmentation network trained with Dice loss and augmentation, and radiomics-based survival regression. The architecture aggregates contextual features, precisely localizes tumor structures, and uses deep supervision to form its output.
- 2 Methods: MRI modalities are normalized across patients and scanners so their value ranges match before neural-network processing.The preprocessing targets both between-patient and between-modality consistency.
- 2.1 Segmentation: The network processes 128x128x128 voxel 3D blocks and intrinsically recombines information across spatial scales.This U-Net-based design avoids manually combining separate input resolutions or filter-size pathways.
- 2.1 Segmentation: The context pathway encodes increasingly abstract features, while the localization pathway combines them with shallower features to localize structures precisely.Higher levels represent lower spatial resolution and higher-dimensional features.
- 2.1 Segmentation: Context modules use pre-activation residual blocks with 3x3x3 convolutions and dropout, connected by stride-2 convolutions for downsampling.Each context module places dropout with pdrop = 0.3 between its two convolutional layers.
- 2.1 Segmentation: The localization pathway upsamples by voxel repetition, applies a 3x3x3 convolution, and uses deep supervision to sum segmentation outputs from multiple levels.The upsampling convolution halves the number of feature maps.
- 2.1 Segmentation: Training uses randomly sampled 128x128x128 patches with batch size 2 and a multiclass Dice loss to address severe class imbalance.The training data contain 166 times as much background as enhancing tumor.
- 2.1 Segmentation: Extensive on-the-fly augmentation uses aggressive parameters that are attenuated during training to prevent overfitting.The passage lists random rotations, scaling, elastic deformations, and gamma correction among the applied techniques.
- 2.2 Survival Prediction: Survival prediction uses radiomic shape, first-order, and gray-level co-occurrence features computed for five tumor regions and fed to a regression ensemble.The ensemble combines a random forest regressor with a multilayer perceptron, using 518 features and 163 training cases.
3 Results
The study reports segmentation and survival-prediction results on BraTS datasets, including benchmark comparisons, qualitative error analysis, and test-set performance. The segmentation model performed strongly overall, while the main tumor-core error involved non-enhancing tumor regions.
- The model ranked first on the BraTS 2015 test-set online leaderboard against other state-of-the-art methods.
- Qualitative results showed accurate segmentation of large tumor regions and fine-grained details, including an enhancing-region wall identified with voxel-level accuracy.
- The most prominent tumor-core error involved non-enhancing tumor, whose poorly defined training labels led the algorithm to rely on context and sometimes guess placement.
- Dice scores on the BraTS 2017 validation set were 0.896 for whole tumor, 0.797 for tumor core, and 0.732 for enhancing tumor.
- Test-set mean Dice scores were 0.858 for whole tumor, 0.775 for tumor core, and 0.647 for enhancing tumor.
- The survival-prediction ensemble achieved 457.83 RMSE, MSE 209607, 52.6% accuracy, and a Spearman correlation coefficient of 0.496 on the test set.
4 Discussion
The paper reports strong segmentation performance on BraTS test data using a U-Net-inspired network, while noting limited exploration of architectural and augmentation variants. Its survival-prediction ensemble achieved moderate test-set performance using shape-based features, with further improvement suggested through additional anatomical information.
- Segmentation: 0.858, 0.775 and 0.647 mean Dice scores were obtained for whole tumor, tumor core and contrast-enhancing tumor on the test set.
- Segmentation: The segmentation network was trained from scratch using provided training data, extensive augmentation and a dice loss formulation.
- Limitations and future work: Training time was about five days per network, and time restrictions limited the number of architectural variants and augmentation methods explored.
- Survival prediction: 52.6% accuracy, 0.496 Spearman correlation and 209607 mean square error were achieved for survival prediction on the test set.
- Survival prediction: The survival approach used an ensemble of a random forest regressor and multilayer perceptrons with shape-based features, while tumor position relative to brain structures was identified as a possible improvement.