Source-linked AI summary
Conditional Adversarial Network for Semantic Segmentation of Brain Tumor
Mina Rezaei, Konstantin Harmuth, Willi Gierke, Thomas Kellermeier, Martin Fischer, Haojin Yang, Christoph Meinel
TL;DR
MRI acquisition differences make brain-tumor segmentation challenging, while the paper addresses both segmentation and survival-day prediction. It proposes conditional-adversarial segmentation and a trainable CNN approach to survival prediction. The segmentation approach reports whole-tumor Dice of 0.68, Specificity of 0.99, and Sensitivity of 0.98.
Problem
MRI acquisition settings, geometry, and hardware vary, changing tumor appearance and making automated brain-lesion segmentation challenging.
Source: §1 · p. 1Method
The paper proposes end-to-end conditional-adversarial tumor segmentation and a trainable survival-day prediction model using clinical data and MR images.
Source: §1 · p. 2, §2.2 · p. 4Results
Takeaways & Limitations
The conclusion presents task-adapted network losses as making the approaches applicable to unseen data.
Source: §4 · p. 8
Abstract
from arXiv · showhide
Automated medical image analysis has a significant value in diagnosis and treatment of lesions. Brain tumors segmentation has a special importance and difficulty due to the difference in appearances and shapes of the different tumor regions in magnetic resonance images. Additionally, the data sets are heterogeneous and usually limited in size in comparison with the computer vision problems. The recently proposed adversarial training has shown promising results in generative image modeling. In this paper, we propose a novel end-to-end trainable architecture for brain tumor semantic segmentation through conditional adversarial training. We exploit conditional Generative Adversarial Network (cGAN) and train a semantic segmentation Convolution Neural Network (CNN) along with an adversarial network that discriminates segmentation maps coming from the ground truth or from the segmentation network for BraTS 2017 segmentation task[15, 4, 2, 3]. We also propose an end-to-end trainable CNN for survival day prediction based on deep learning techniques for BraTS 2017 prediction task [15, 4, 2, 3]. The experimental results demonstrate the superior ability of the proposed approach for both tasks. The proposed model achieves on validation data a DICE score, Sensitivity and Specificity respectively 0.68, 0.99 and 0.98 for the whole tumor, regarding online judgment system.
1 Introduction
The introduction frames brain-tumor segmentation as challenging because MRI acquisition and hardware differences alter tumor appearance. It presents conditional-adversarial segmentation and a separate survival-day prediction model for BraTS 2017.
1 Introduction
- MRI acquisition settings, geometry, and hardware vary, changing tumor appearance and making automated segmentation challenging.Source: §1 · p. 1
- The paper addresses BraTS-2017 segmentation with an end-to-end adversarial network and survival-day prediction with a trainable model based on clinical data and MR images.The segmentation target is described as High and Low Grade Glioma tumors.Source: §1 · p. 2
- 0.68 overall Dice for whole-tumor, 0.99 Specificity, and 0.98 Sensitivity are reported for the proposed segmentation approach.The introduction also describes the segmentation result as covering two types of brain tumor.Source: §1 · p. 2
2 Methodology
The methodology describes a conditional GAN approach for multi-class tumor segmentation and a separate two-path CNN approach for survival-day prediction. The segmentation model uses a U-Net generator and a Markovian GAN discriminator.
2 Methodology
- The proposed cGAN approach places the method among GANs with multiple output classes, using three segmentation classes for each multimodal image.The methodology first frames the discriminator as distinguishing reference-distribution samples from generated outputs.Source: §2 · pp. 2–3
2.1 Brain Tumor Semantic Segmentation
- The segmentation model uses a patient-wise U-Net generator and a Markovian GAN discriminator trained simultaneously through backpropagation.Virtual-BatchNorm is applied to the generator and Reference-BatchNorm to the discriminator to reduce over-fitting; the U-Net is chosen to retain whole-image context.Source: §2.1 · pp. 3–4
2.2 Survival Day Prediction
- The survival-day model has separate CNN paths for multimodal MRI and clinical features, concatenating their representations before two fully connected layers.The image paths use Virtual-BatchNorm; training uses image flips and rescaling for augmentation and mean squared error as the loss.Source: §2.2 · p. 4
3 Experiments
The experiments evaluate cGAN segmentation on BraTS 2017 and report preliminary online-validation results, alongside a separate survival-day prediction evaluation.
3 Experiments
- The cGAN evaluation uses BraTS 2017 data, bias-field correction and histogram matching, with 75 LGG and 210 HGG patients in training.Training used all released challenge data and took around 72 hours on parallel Pascal Titan X GPUs.Source: §3 · p. 4
- Inputs comprise four MRI modalities across x, y, and z axes, and outputs comprise three tumor sub-regions across 530,100 2D images.The passage reports 3×4×155×285 prepared images and 250×250 image size.Source: §3 · p. 4
- Segmentation evaluation uses Dice score, Hausdorff distance, sensitivity, and specificity for whole, core, and enhancing tumor regions.The regions are defined as WT, CT, and ET, respectively.Source: §3 · p. 6
- BraTS-2017 online-judge results on unseen validation data are described as preliminary and still in progress.Table 1 reports the segmentation results.Source: §3 · p. 6, Table 1 · p. 6
- Survival-day prediction results are reported separately in Table 2, using 115 training, 16 validation, and 32 testing patients.The first CNN pathway has seven input channels: four from multimodal images and three from segmented regions.Source: Table 2 · p. 6
4 Conclusion
The paper presents approaches for brain-tumor segmentation and survival-day prediction. It describes adversarial training for segmentation and a parallel CNN for survival prediction, with future work targeting an RNN in the segmentation network.
4 Conclusion
- The segmentation approach is an end-to-end trainable conditional GAN, with adversarial training used to reduce over-fitting and increase robustness.The conclusion describes the generator CNN as global-based.Source: §4 · p. 8
- A separate automated, trainable parallel CNN predicts survival day after tumor diagnosis.Source: §4 · p. 8
- The networks learn losses adapted to the task and data, which the authors say makes them applicable to unseen data.Source: §4 · p. 8
- Future work will explore incorporating an RNN inside the segmentation network’s encoder-decoder.Source: §4 · p. 8