Source-linked AI summary
Automatic Brain Tumor Segmentation using Cascaded Anisotropic Convolutional Neural Networks
Guotai Wang, Wenqi Li, Sebastien Ourselin, Tom Vercauteren
TL;DR
Automatic brain-tumor segmentation is challenging because tumors vary substantially across patients and have ambiguous boundaries. The paper proposes a hierarchical cascade of binary CNN segmenters with anisotropic processing and multi-view fusion, achieving reported Dice scores across BraTS 2017 validation and testing sets. Its discussion also identifies weighted view averaging and CRF post-processing as possible directions beyond the current results.
Problem
Automatic brain-tumor segmentation is difficult because tumor size, shape, and location vary across patients and boundaries can be ambiguous.
Method
The method hierarchically segments whole tumor, tumor core, and enhancing tumor core with cascaded binary CNNs using anisotropic networks, multi-scale and residual features, and three-view fusion.
Results
The method achieved average Dice scores of 0.7859, 0.9050, and 0.8378 on BraTS 2017 validation, and 0.7831, 0.8739, and 0.7748 on testing, for enhancing tumor core, whole tumor, and tumor core, respectively.
Takeaways & Limitations
The proposed cascade provides competitive automatic segmentation of glioma subregions from multi-modal brain MR images.
Takeaways & Limitations
The current testing-stage fusion uses a simple average despite differing imaging resolutions across views, and the results do not use CRF post-processing.
Abstract
from arXiv · showhide
A cascade of fully convolutional neural networks is proposed to segment multi-modal Magnetic Resonance (MR) images with brain tumor into background and three hierarchical regions: whole tumor, tumor core and enhancing tumor core. The cascade is designed to decompose the multi-class segmentation problem into a sequence of three binary segmentation problems according to the subregion hierarchy. The whole tumor is segmented in the first step and the bounding box of the result is used for the tumor core segmentation in the second step. The enhancing tumor core is then segmented based on the bounding box of the tumor core segmentation result. Our networks consist of multiple layers of anisotropic and dilated convolution filters, and they are combined with multi-view fusion to reduce false positives. Residual connections and multi-scale predictions are employed in these networks to boost the segmentation performance. Experiments with BraTS 2017 validation set show that the proposed method achieved average Dice scores of 0.7859, 0.9050, 0.8378 for enhancing tumor core, whole tumor and tumor core, respectively. The corresponding values for BraTS 2017 testing set were 0.7831, 0.8739, and 0.7748, respectively.
1 Introduction
Brain-tumor segmentation can support reproducible tumor measurement, but variation across patients and ambiguous boundaries make automatic segmentation difficult. The paper addresses this with a hierarchical CNN cascade and anisotropic, multi-view design choices.
- Automatic segmentation could provide accurate, reproducible tumor measurements for diagnosis, surgical planning, and treatment assessment.
- Tumor size, shape, and location vary considerably across patients, limiting the usefulness of prior anatomical shape and-location information.
- Ambiguous boundaries arise from smooth intensity gradients, partial-volume effects, and bias-field artifacts.
- The proposed cascade decomposes multi-class segmentation into three simpler binary problems using the hierarchical structure of tumor subregions.
- Anisotropic convolution balances receptive field, model complexity, and memory consumption, while dilated convolution, residual connections, and multi-scale prediction target segmentation performance.
- Multi-view fusion combines predictions from three orthogonal views for more robust brain-tumor segmentation.
2 Methods
The method uses three fully convolutional networks in a hierarchical cascade, combining anisotropic 3D processing with residual, multi-scale, and orthogonal-view fusion. Bounding-box localization progressively narrows the regions segmented by later networks.
- Triple Cascaded Framework: WNet segments the whole tumor, TNet segments the tumor core within WNet’s bounding box, and ENet segments enhancing tumor core within TNet’s bounding box.
- Anisotropic Convolutional Neural Networks: Anisotropic networks use slice stacks with large in-plane and relatively small out-of-plane receptive fields to balance context, complexity, and memory.The 2D receptive fields are 217×217 for WNet and TNet and 113×113 for ENet; all use an out-of-plane receptive field of 9.
- Anisotropic Convolutional Neural Networks: A 3×3×3 kernel is decomposed into 3×3×1 intra-slice and 1×1×3 inter-slice kernels, with batch normalization and PReLU activation following each convolution.
- Residual Connection: Residual blocks add each block’s input directly to its output to support information propagation and speed convergence.WNet, TNet, and ENet each use 10 residual blocks.
- Multi-scale Prediction: Multi-scale prediction upsamples intermediate score maps, concatenates them, and applies another convolution to produce the final score map.Each network predicts two segmentation classes.
- Multi-view Fusion: Predictions from axial, sagittal, and coronal networks are averaged at each cascade level to incorporate 3D contextual information.Multi-view fusion is applied to WNet, TNet, and ENet during testing.
3 Experiments and Results
Experiments on BraTS 2017 evaluated the cascaded networks and multi-view fusion using multimodal, co-registered MR images. The proposed method achieved strong Dice scores across tumor subregions on validation and testing sets.
- Dataset: BraTS 2017 experiments used four co-registered MR sequences from 285 training, 46 validation, and 146 testing patients.The sequences were T1, T1c, T2, and FLAIR; images were skull-striped and resampled to 1mm3 isotropic resolution.
- Qualitative results: Multi-view fusion produced more accurate tumor-core segmentation than axial-view networks alone, while whole-tumor results were similar.The comparison used training-image examples from both HGG and LGG cases.
- Quantitative results: 0.7859, 0.9050, and 0.8378 were the proposed method’s average Dice scores for enhancing tumor core, whole tumor, and tumor core on validation data.The corresponding no-fusion scores were 0.7411, 0.8896, and 0.8255, respectively.
- Quantitative results: 0.7831, 0.8739, and 0.7748 were the proposed method’s average Dice scores for enhancing tumor core, whole tumor, and tumor core on testing data.Table 2 also reported standard deviations, medians, and 25 and 75 quantiles for Dice and Hausdorff distance.
- Quantitative results: Higher median values indicated good segmentation for most testing images, while some outliers lowered the average scores.This interpretation was reported alongside the testing-set evaluation.
4 Discussion and Conclusion
The discussion attributes the method’s performance to hierarchical cascading, anisotropic network design, and three-view fusion. It also identifies longer training and testing time and possible future refinement with weighted fusion or CRFs.
- Discussion: The cascade simplifies hierarchical tumor segmentation into three binary networks and uses spatial constraints to reduce false positives.Tumor core is restricted inside whole tumor, and enhancing tumor core inside tumor core.
- Limitations: The cascade is not end-to-end and requires longer training and testing than its multi-class alternative.This is identified as a drawback of the cascaded framework.
- Discussion: Three-view fusion uses axial, sagittal, and coronal predictions to incorporate 3D contextual information and improve segmentation accuracy.The current fusion averages outputs, while weighted averaging is suggested as potentially more reasonable because imaging resolution differs across views.
- Limitations: The current results omit CRF post-processing, leaving spatial regularization as a possible route for further improvement.The discussion states that CRFs have been effective for obtaining more spatially regularized segmentation.
- Discussion: Anisotropic receptive fields balance receptive field size, model complexity, and memory consumption for 3D MR images.The networks combine this design with dilated convolution, residual connections, and multi-scale prediction.