Source-linked AI summary
3D Dilated Multi-Fiber Network for Real-time Brain Tumor Segmentation in MRI
Chen Chen, Xiaopeng Liu, Meng Ding, Junfeng Zheng, Jiangyun Li
TL;DR
Accurate brain tumor segmentation in MRI is computationally demanding because 3D CNNs process volumetric context with substantial overhead. The paper proposes DMFNet, combining lightweight multi-fiber units with weighted 3D dilated convolutions for efficient multi-scale segmentation. On BraTS 2018, it reports comparable accuracy with substantially lower parameter and FLOP counts, including real-time inference.
Problem
3D CNNs capture volumetric structure for brain tumor segmentation but incur high computational cost, while existing efficient models do not match state-of-the-art performance.
Method
DMFNet combines efficient multi-fiber group convolution with learnable weighted 3D dilated convolution to reduce computation and represent multi-scale features.
Results
80.12%, 90.62% and 84.54% Dice scores for ET, WT and TC were achieved with 10× less model parameters and 50× less computational FLOPs than the state-of-the-art algorithm.
Takeaways & Limitations
DMFNet provides a lightweight approach for real-time 3D MRI brain tumor segmentation while maintaining comparable segmentation accuracy.
Abstract
from arXiv · showhide
Brain tumor segmentation plays a pivotal role in medical image processing. In this work, we aim to segment brain MRI volumes. 3D convolution neural networks (CNN) such as 3D U-Net and V-Net employing 3D convolutions to capture the correlation between adjacent slices have achieved impressive segmentation results. However, these 3D CNN architectures come with high computational overheads due to multiple layers of 3D convolutions, which may make these models prohibitive for practical large-scale applications. To this end, we propose a highly efficient 3D CNN to achieve real-time dense volumetric segmentation. The network leverages the 3D multi-fiber unit which consists of an ensemble of lightweight 3D convolutional networks to significantly reduce the computational cost. Moreover, 3D dilated convolutions are used to build multi-scale feature representations. Extensive experimental results on the BraTS-2018 challenge dataset show that the proposed architecture greatly reduces computation cost while maintaining high accuracy for brain tumor segmentation. The source code can be found at https://github.com/China-LiuXiaopeng/BraTS-DMFNet
1 Introduction
Brain tumor segmentation is important for MRI-based diagnosis, monitoring, and treatment planning, but heterogeneous glioma appearance and the computational cost of 3D CNNs make accurate, efficient automation difficult.
- Accurate brain tumor segmentation supports diagnosis and treatment planning, while MRI also informs tumor monitoring and patient outcome prediction.
- Gliomas’ heterogeneous appearance and shape make automated segmentation in multi-modal MRI challenging.
- 3D convolutions capture volumetric structure effectively but incur high computational cost because of the extra dimension.
- Existing lightweight models such as 3D-ESPNet and SD-UNet reduce computational demands, but their performance is not comparable to the state of the art.
- DMFNet addresses the efficiency–accuracy gap using efficient group convolution and weighted 3D dilated convolution for multi-scale representation.
2 Method
DMFNet combines grouped multi-fiber processing, multiplexer-based information routing, and adaptively weighted dilated branches to reduce computation while representing multi-scale 3D context.
- Channel Grouping: Channel grouping splits convolutional channels into parallel fibers, reducing the grouped residual unit’s parameter amount to 1/g of the regular unit.The formulation uses a constant 3 × 3 × 3 kernel and g parallel residual units.
- Multiplexer: The multiplexer uses two 1 × 1 × 1 convolutions to route information among fibers while squeezing channels to c_in/4 and restoring them to c_in.Residual shortcuts outside the multiplexer and unit pass information directly from lower to higher levels without additional parameters.
- Dilated Fiber: The dilated fiber uses three 3D convolution branches with dilation rates d = 1, 2, and 3, whose learnable weighted sum captures multi-scale spatial correlations.The weights are assigned to the branches and summed to select information from different fields of view.
- DMFNet Architecture: DMFNet uses MF/DMF units as encoder-decoder building blocks, applying DMF units in the first six encoding units for multi-scale representation.The network receives four-channel input corresponding to four-modal MRI data and uses 2-stride convolution for downsampling.
3 Experiments and Results
DMFNet achieves comparable BraTS 2018 segmentation accuracy with substantially lower model complexity and fast inference. Its dilated convolutions and adaptive weighting improve multi-scale representation and segmentation performance.
- Dataset and setup: 285 training and 66 validation cases comprise the BraTS 2018 dataset, with four MRI sequences and three tumor-region labels.Each volume is 240×240×155.
- Segmentation results: 80.12%, 90.62% and 84.54% are DMFNet’s validation Dice scores for ET, WT and TC, respectively.Compared with NVDLMED, the performance gaps are 1.61% for ET, 0.06% for WT and 1.48% for TC.
- Model efficiency: 10× fewer parameters and 55× fewer FLOPs allow DMFNet to retain comparable accuracy to the best single-model baseline.The paper also reports average inference times of 0.019s on one GPU and 20.6s on one CPU per 3D MR image.
- Model efficiency: 1.81M parameters and 13.36G FLOPs are achieved by the 0.75× MFNet variant while retaining strong Dice scores.This variant reduces MFNet’s channel count to 75%.
- Ablation analysis: Adaptive weighting of dilated branches produces more favorable scores than equal weighting, particularly for ET.The ablation compares learnable weights with ω1 = ω2 = ω3 = 1.
- Ablation analysis: ω1 dominates the first DMF unit, while ω3 increasingly influences DMF units 2–6 in higher layers.ω1 corresponds to a small receptive field and ω3 to a large receptive field.
- Qualitative results: Red, green and blue regions in the visual comparison denote necrotic or non-enhancing tumor, peritumoral edema and GD-enhancing tumor.GT denotes the ground truth.
4 Conclusion
The proposed Dilated Multi-Fiber network combines multi-fiber units with learnable weighted 3D dilated convolutions for efficient 3D MRI tumor segmentation. It reports comparable Dice scores with substantially lower computational cost and supports real-time inference.
- Conclusion: 3.88M parameters and around 27G FLOPs enable DMFNet to achieve real-time inference for 3D brain tumor segmentation.The network uses multi-fiber units to reduce 3D convolution cost and weighted dilated convolutions for multi-scale representation.
- Conclusion: 80.12%, 90.62% and 84.54% are the reported ET, WT and TC Dice scores on the BraTS 2018 challenge.These scores are reported alongside 10× fewer parameters and 50× less computational FLOPs than NVDLMED.
- Conclusion: Table 2 evaluates the effect of the weighting strategy for the dilated sub-fibers.The supplied conclusion materials identify the weighting strategy as part of the proposed design.