Abstract
Accurate automatic liver and tumor segmentation plays a vital role in treatment planning and disease monitoring. Recently, deep convolutional neural networks (DCNNs) have obtained tremendous success in 2D and 3D medical image segmentation. However, 2D DCNNs cannot fully leverage inter-slice information, while 3D DCNNs are computationally expensive and memory intensive. To address these issues, we propose a novel dense-sparse training flow from a data perspective, in which densely adjacent slices and sparsely adjacent slices are extracted as inputs for regularizing DCNNs. Moreover, we design a 2.5D lightweight nnU-Net using depthwise separable convolutions to improve efficiency. Extensive experiments on the LiTS dataset demonstrate the superiority of the proposed method, achieving high accuracy with significantly reduced complexity.
Introduction
In medical imaging, particularly CT scans, precise segmentation of organs and pathologies is critical for clinical workflows. While 2D CNNs are computationally efficient, they lack 3D context, often leading to inconsistent results across slices. Conversely, 3D CNNs capture volumetric context but demand excessive memory and computation, making them difficult to deploy on standard clinical hardware. This research introduces a "2.5D" approach that bridges this gap, providing 3D spatial awareness with the efficiency of 2D architectures.
Methodology
The proposed framework addresses efficiency and performance through two primary innovations: data sampling and network architecture.
Dense-Sparse (DS) Sampling
Instead of standard consecutive slice feeding, the model utilizes two input types to regularize learning:
- Densely adjacent slices (stride = 1): Provide fine-grained local details.
- Sparsely adjacent slices (stride = 2): Offer a wider field of view along the depth axis, helping the model learn multi-scale spatial features.
DS nnU-Net Architecture
The network is built upon the popular nnU-Net framework but incorporates Depthwise Separable Convolutions. This modification splits standard convolutions into depthwise and pointwise steps, drastically reducing the parameter count and floating-point operations (FLOPs) without sacrificing the ability to learn complex features.
Training Strategy
A two-stage Dense-Sparse-Dense (DSD) training strategy is employed. Initial training uses a mix of dense and sparse samples for broad feature learning and fast convergence, followed by fine-tuning on dense samples to ensure robustness and sharp segmentation boundaries.
Experimental Results
The model was evaluated on the LiTS (Liver Tumor Segmentation) dataset, which includes 201 CT scans. Preprocessing included resizing to $512 \times 512$, resampling, and normalization to ensure consistency across different scanners.
Performance and Efficiency
The proposed DS nnU-Net-DSD achieved a Dice score of 0.814 for lesions and 0.962 for the liver. These results are nearly identical to the performance of the much heavier 3D nnU-Net baseline, as shown in the comparative data below.
The most significant achievement is the reduction in model size: the DS nnU-Net-DSD uses only 7 million parameters, compared to 37-41 million for standard 2D and 3D baselines. This efficiency translates to a 25% reduction in training time on consumer-grade hardware, making it highly accessible for clinical practice.
Conclusion
This work demonstrates that 3D context can be effectively captured using a 2.5D approach combined with smart data sampling. By reducing the parameter count by nearly 6x while maintaining competitive accuracy, this method offers a practical solution for efficient medical image segmentation. Future research will explore the generalization of this framework to other organs and imaging modalities like MRI.
Summary updated