Source-linked AI summary
Cascaded V-Net using ROI masks for brain tumor segmentation
Adrià Casamitjana, Marcel Catà, Irina Sánchez, Marc Combalia, Verónica Vilaplana
TL;DR
Brain tumor MRI segmentation is challenging because tumors are heterogeneous, have variable fuzzy boundaries, and occupy a small fraction of the image. The paper uses a cascade of modified V-Net CNNs with ROI masks for whole-tumor localization and sub-region segmentation, achieving competitive whole-tumor Dice performance but weaker enhancement and tumor-core delineation on BraTS2017.
Problem
Brain tumor segmentation is difficult because gliomas contain heterogeneous regions, tumor boundaries vary across patients, and tumors occupy only a small portion of brain images.
Method
A cascade of two modified V-Net CNNs uses brain and tumor-vicinity ROI masks to localize whole tumors and then segment sub-regions with dense training.
Results
The method achieves competitive whole-tumor Dice performance on BraTS2017 but low Dice for enhancing-tumor and tumor-core regions.
Takeaways & Limitations
ROI-constrained cascaded dense training can produce competitive whole-tumor segmentation while remaining insufficient for reliably capturing less common tumor regions.
Takeaways & Limitations
Enhancing-tumor and tumor-core regions are underrepresented in predictions, and the authors identify this behavior as a target for future improvement.
Abstract
from arXiv · showhide
In this work we approach the brain tumor segmentation problem with a cascade of two CNNs inspired in the V-Net architecture \cite{VNet}, reformulating residual connections and making use of ROI masks to constrain the networks to train only on relevant voxels. This architecture allows dense training on problems with highly skewed class distributions, such as brain tumor segmentation, by focusing training only on the vecinity of the tumor area. We report results on BraTS2017 Training and Validation sets.
1 Introduction
Brain tumor segmentation remains difficult because gliomas are heterogeneous and tumor boundaries vary across patients, while manual annotation is time-consuming. The paper addresses this challenge with deep CNNs and ROI-constrained cascaded training focused on relevant voxels.
- Brain MRI tumor segmentation supports progression monitoring, treatment planning, and follow-up, but manual segmentation is time-consuming and subject to inter-rater discrepancy.
- Gliomas contain heterogeneous sub-regions with differing MRI intensity profiles, and tumor borders vary in shape, location, and extent across patients.
- Deep learning models learn increasingly complex task-specific features directly from data rather than relying on predefined features.
- Some prior approaches process 2D slices or orthogonal patches, while fully 3D methods can use parallel pathways to capture local and global context.
- The proposed method cascades two modified V-Net CNNs and uses ROI masks so each network focuses on relevant voxels or tumor vicinity.
- ROI-constrained dense training computes loss only within relevant masks, blocking backpropagated signals from outer voxels and avoiding patch-wise training.
2 Method
The method divides brain tumor segmentation into whole-tumor localization followed by sub-region delineation using a cascade of modified V-Net networks. ROI masks constrain training and inference to brain tissue or tumor vicinity, supporting dense training under severe class imbalance.
- 2 Method: Brain tumors occupy a small fraction of images, so the method separates whole-tumor segmentation from delineation of individual tumor regions.The two CNNs operate as linked stages, with the first network’s output supplied to the second.
- 2.1 V-Net using ROI masks: The networks use modified V-Net blocks with 3x3x3 filters, ReLU activations, batch normalization, and reformulated identity residual connections.Dimension mismatches are handled with pooling, up-sampling, and 1x1x1 convolutions.
- 2.1 V-Net using ROI masks: ROI masks force outer voxels to background during training and inference, preventing their backpropagated signals from contributing to learning.
- 2.2 Training: Dense training uses one subject per batch instead of patch-wise training or non-uniform sampling strategies for class imbalance.The first network performs binary whole-tumor segmentation from four MRI modalities and uses a brain mask during training.
- 2.2 Training: The modified Dice coefficient is used for binary segmentation with imbalanced data; p_i denotes voxel softmax output and l_i its binary label.
- 2.2 Training: The second network performs four-class segmentation near the tumor using a rectangular whole-tumor mask and a combined cross-entropy plus sub-region Dice loss.Its classes are non-tumor, edema, enhancing core, and non-enhancing core.
- 2.2 Training: During inference, morphological filtering cleans the first prediction and its smallest enclosing rectangle becomes the second network’s ROI mask.This masks most false positives outside the tumor vicinity.
3 Results and discussion
On BraTS2017, the cascaded V-Net correctly segmented whole tumors but was weaker for smaller enhancing and tumor-core regions. Training curves and visual examples indicate overfitting and class underrepresentation, especially for these smaller regions.
- Data: BraTS2017 contains 285 training scans with four MRI modalities and annotations for enhancing tumor, edema, and necrotic or non-enhancing tumor.The data includes 210 glioblastoma or high-grade glioma scans and 75 lower-grade glioma scans.
- Performance evaluation: The evaluation reports Dice, Hausdorff distance, sensitivity, and specificity for whole tumor, tumor core, and enhancing tumor.Results are reported for a 70% training and 30% development split and for the BraTS2017 Validation leaderboard.
- Quantitative results: The method achieves high or competitive whole-tumor Dice, but relatively low enhancing-tumor and tumor-core Dice compared with state-of-the-art methods.On the Validation set, whole-tumor Dice is close to top-performing participants, while enhancing-tumor and tumor-core Dice remain low.
- Quantitative results: Sensitivity suggests that enhancing-tumor and tumor-core regions are underrepresented in predictions, despite proper background representation associated with prediction masks.Specificity is less informative for imbalanced classes, whereas sensitivity exposes the weaker detection of smaller regions.
- Training curves: From epoch 18, tumor-core performance begins overfitting, while whole-tumor improvement is insignificant and enhancing-tumor performance stops improving despite continued development-loss reduction.The continuing loss improvement is attributed to the cross-entropy term.
- Visual analysis: Visual examples show correct whole-tumor localization, but edema can be overrepresented at the expense of smaller enhancing and necrotic or non-enhancing regions.One example captures all tumor regions, whereas another localizes the tumor but misses distinctions among subregions.
4 Conclusions
The paper concludes that masked cascaded V-Net enables dense training focused on relevant brain regions, while smaller tumor classes remain difficult to detect.
- 4 Conclusions: The cascaded V-Net uses masks to focus training on relevant brain regions and addresses class imbalance through separate tumor-localization and subregion-segmentation steps.The approach performs well for whole-tumor segmentation but does not properly capture less common tumor-core and enhancing-tumor regions.
- 4 Conclusions: Smaller tumor regions remain underdetected even with ROI masks, motivating future work that increases their learning weight.The stated future direction is to up-weight small tumor regions during learning.