Source-linked AI summary
MRI Tumor Segmentation with Densely Connected 3D CNN
Lele Chen, Yue Wu, Adora M. DSouza, Anas Z. Abidin, Axel Wismuller, Chenliang Xu
TL;DR
Small glioma lesions can make MRI segmentation vulnerable to class imbalance, despite its value for treatment planning and disease monitoring. The paper combines patch-wise training with a multiscale, hierarchical, densely connected 3D CNN; on BraTS 2017 it reports Dice scores of 0.72, 0.83, and 0.81 for complete tumor, tumor core, and enhancing tumor, respectively.
Problem
Small lesion areas in MRI scans create a class-imbalance challenge for brain tumor segmentation, where background can dominate training.
Source: §1 · p. 1Method
The model uses patch-wise training and a multistage 3D CNN with two receptive-field scales, dense blocks, and hierarchical segmentation of lesion regions.
Source: §1 · pp. 1–2Results
0.72, 0.83, and 0.81 Dice scores for complete tumor, tumor core, and enhancing tumor segmentation, respectively, were comparable to state-of-the-art results.
Source: §1 · p. 2, §3.3.3 · p. 5Takeaways & Limitations
The results suggest potential for lesion segmentation tasks in MRI scans.
Source: §1 · p. 2In observed examples, predictions were inconsistent along the axis perpendicular to the axial plane.
Source: §3.3.2 · p. 5
Abstract
from arXiv · showhide
Glioma is one of the most common and aggressive types of primary brain tumors. The accurate segmentation of subcortical brain structures is crucial to the study of gliomas in that it helps the monitoring of the progression of gliomas and aids the evaluation of treatment outcomes. However, the large amount of required human labor makes it difficult to obtain the manually segmented Magnetic Resonance Imaging (MRI) data, limiting the use of precise quantitative measurements in the clinical practice. In this work, we try to address this problem by developing a 3D Convolutional Neural Network~(3D CNN) based model to automatically segment gliomas. The major difficulty of our segmentation model comes with the fact that the location, structure, and shape of gliomas vary significantly among different patients. In order to accurately classify each voxel, our model captures multi-scale contextual information by extracting features from two scales of receptive fields. To fully exploit the tumor structure, we propose a novel architecture that hierarchically segments different lesion regions of the necrotic and non-enhancing tumor~(NCR/NET), peritumoral edema~(ED) and GD-enhancing tumor~(ET). Additionally, we utilize densely connected convolutional blocks to further boost the performance. We train our model with a patch-wise training schema to mitigate the class imbalance problem. The proposed method is validated on the BraTS 2017 dataset and it achieves Dice scores of 0.72, 0.83 and 0.81 for the complete tumor, tumor core and enhancing tumor, respectively. These results are comparable to the reported state-of-the-art results, and our method is better than existing 3D-based methods in terms of compactness, time and space efficiency.
1. INTRODUCTION
Glioma segmentation in MRI is difficult in part because small lesions create class imbalance, while accurate segmentation supports treatment planning and disease monitoring. The paper addresses this with a multistage 3D CNN that combines multiple receptive-field scales, dense connections, and hierarchical lesion segmentation.
1. INTRODUCTION
- Small lesion areas can let background voxels dominate MRI training, motivating patch-wise sampling centered on lesion areas.Patch-wise training is described as a way to balance positive and negative samples.Source: §1 · p. 1
- The multistage model predicts each voxel using two receptive-field sizes and adds classification scores to form the final score.The two scales are intended to capture both nearby and more distant context.Source: §1 · p. 2
- The paper applies densely connected blocks to 3D MRI segmentation, addressing whether dense connections improve brain tumor segmentation.The introduction identifies that question as unexplored and describes the paper's contributions as including 3D dense blocks, multiscale receptive fields, and lesion-structure-inspired hierarchy.Source: §1 · p. 2
- The architecture hierarchically segments edema and its enhancing and necrotic or non-enhancing subregions, treating those subregions as lying inside edema.One pathway segments edema using FLAIR and T2 scans; a second uses all MRI scan types to segment edema and its subregions.Source: §1 · p. 2
- On the BraTS 2017 benchmark, experiments indicate that the model achieved performance comparable to state-of-the-art methods without bells and whistles, suggesting potential for lesion segmentation tasks in MRI scans.Source: §1 · p. 2
2. METHODS
The method uses a hierarchical two-pathway model: one pathway detects whole tumor, and a second classifies lesion regions using combined modality features. Its feature extractors use two densely connected stages with different receptive-field scales.
2.1 Hierarchical Segmentation Structure
- A whole-tumor binary pathway feeds FLAIR/T2 features into a four-class pathway that combines T1/T1-CE features to classify background, ED, ET, and NCR/NET.The pathways are trained together, and final predictions come from the four-class pathway.Source: §2.1 · p. 2
2.2 Two-Stage Densely Connected 3D CNN
- Each of the two feature-extractor stages uses a six-layer densely connected block with growth rate 12.The initial convolution uses 24 kernels.Source: §2.2 · p. 3, Table 1 · p. 4
- Each dense-block layer concatenates outputs from all preceding layers before convolution, with batch normalization and ReLU in the transformation.The post-block 1×1×1 convolution merges features from all abstraction levels.Source: §2.2 · pp. 3–4
- With stride set to 1 and no downsampling layers to preserve high-resolution patches, the two stages have receptive fields of 15 and 27, respectively, in each dimension of a 3D kernel.Source: §2.2 · p. 4
3. EXPERIMENT AND RESULTS
On BraTS 2017, the proposed model achieves its highest Dice scores for tumor core and enhancing tumor segmentation, while experiments examine its data, training, and architectural components. Qualitative examples also reveal inconsistent predictions across slices perpendicular to the axial plane.
3.1 Data
- The test set contains 42 of 210 BraTS 2017 sequences; four MRI modalities provide complementary information for segmenting whole tumors and subregions.N4ITK bias correction and per-sequence mean and standard-deviation normalization are applied to the scans.Source: §3.1 · p. 4
3.2 Training and Testing
- Training uses 3D patches sampled with equal numbers centered on lesion and background; four-fold validation tunes hyperparameters because there is no separate validation set.Testing uses non-overlapping patches and a sliding window to obtain dense predictions.Source: §3.1 · p. 4, §3.2 · pp. 4–5
3.3.1 Metric
- Dice evaluates overlap between predicted lesion regions and ground truth, with true-positive, false-positive, and false-negative predictions defining the measure.The paper reports testing results using Dice.Source: §3.3.1 · p. 5
3.3.2 Qualitative Results
- In two selected MRI sequences, axial examples show promising subregion segmentation, but predictions can be inconsistent across the perpendicular coronal and sagittal views.The authors suggest patch-wise training may limit capture of long-term dependencies in MRI sequences.Source: §3.3.2 · p. 5, Figure 3 · p. 6
3.3.3 Quantitative Results
- The method scored Dice 0.72, 0.83, and 0.81 for complete tumor, tumor core, and enhancing tumor segmentation, respectively; it led for tumor core and enhancing tumor, while Kamnitsas et al.3 led for complete tumor.Source: §3.3.3 · p. 5
3.3.4 Ablation Study
- The non-hierarchical model comes closest overall, but the proposed model outperforms it by a large margin for tumor core and enhancing tumor; the non-dense model trails in all three categories.The single-scale receptive-field model has the lowest accuracy; the non-dense network also introduces more parameters.Source: §3.3.4 · p. 7, Table 3 · p. 7
4. DISCUSSION AND CONCLUSION
The paper introduces a 3D MRI brain-tumor segmentation approach combining dense connectivity, multi-scale receptive fields, and hierarchical prediction. Experiments report performance on par with state-of-the-art models and support the effectiveness of the architecture's components.
4. DISCUSSION AND CONCLUSION
- The model combines densely connected blocks, multi-scale receptive fields, and hierarchical segmentation for brain tumor MRI.It predicts 123 voxels in one iteration to increase efficiency.Source: §4 · p. 8
- Experimental results indicate that the proposed model performs on par with state-of-the-art models without advanced tricks.Source: §4 · p. 8
- The ablation study supports the effectiveness of each component; the authors conclude the approach has potential for MRI and other medical image segmentation tasks.Source: §4 · p. 8