This document presents a novel deep learning framework for efficient liver and tumor segmentation using a multi-slice dense-sparse approach. It highlights the challenges of traditional 2D and 3D networks and introduces a lightweight depthwise separable NNU-Net that offers a significant reduction in parameters while improving accuracy. Extensive experiments validate the effectiveness of the proposed method in both segmentation accuracy and processing efficiency.

1. Multi-Slice Dense-Sparse Learning for Efficient Liver
and Tumor Segmentation
Ziyuan Zhao1 , Zeyu Ma1,2, Yanjie Liu1,2, Zeng Zeng1, Pierce KH Chow3,4
Presenter: Zhao Ziyuan
1 Institute for Infocomm Research, A*STAR
2 National University of Singapore
3 National Cancer Center and Singapore General Hospital, Singapore
4 Duke-NUS Medical School Singapore

2. Introduction – Liver & Tumor Segmentation
- Liver tumor is one of the main causes of human death.
- Accurate automatic liver and tumor segmentation plays a vital role in treatment planning and disease
monitoring.
- Deep learning has been widely used for medical image segmentation.

3. Challenge – Trade off between Accuracy and Complexity
- 2D DCNNs cannot fully leverage the inter-slice information in 3D volumetric CT scans
- 3D DCNNs are capable to explore the inter-slice correlations and learn deep 3D representations with
volumetric inputs, but they are computationally expensive and memory intensive.
Prediction
Accuracy
Model
Complexity

4. Method – Multi-slice Sampling + Efficient nnU-Net
- Consider both data and network perspectives in our framework for more efficient liver and tumor
segmentation.
- We propose a novel dense-sparse training flow from a data perspective.
- We design a 2.5D light-weight nnU-Net from a network perspective.

5. Method (1) –Dense-Sparse Sampling
- 2.5D network uses a stack of T continuous slices during training and generates the segmentation mask
for the central slice at the inference stage.
- We propose a novel dense-sparse sampling method to generate densely adjacent slices and sparsely
adjacent slices
- Consider different strides of sampling for a more comprehensive view.
s is the stride of sampling
When s = 1, densely adjacent slices
When s > 1, sparsely adjacent slices

6. Method (2) – Depthwise Separable nnU-Net
- Select 2D nnU-Net as our backbone and design a depthwise separable nnU-Net (DS nnU-Net).
- Employ depthwise separable convolutions instead of standard convolution layers to further ease the
computational burden.
- DS nnU-Net has only 7.7 million (M) parameters, while 2D nnU-Net has more than 40 M parameter.

7. Experiments
- Dataset
- Liver Tumor Segmentation (LiTS) dataset
- 201 CT scans (131 for training and 70 for testing)
- 105 volumes for training and the remaining 26 for testing in our experiments
- Data preprocessing
- CT volumes are resized to multiple slices of size 512 × 512 after resampling and normalization
- Implementation details
- Supervised loss: Dice + Cross-entropy
- We implement DS sampling with thickness T = 7.
- For DSD training, we set 400 and 600 epochs for DS step and D step

8. Results (1) – Quantitative Comparison
- Compare our method with 2D nnU-Net and 3D nnUNet with full resolution
- Compare with different variants
- nnU-Net-DS: 2D nnU-Net with the proposed densesparse sampling.
- nnU-Net-DSD: 2D nnU-Net with the proposed dense-sparse-dense training strategy.
- DS nnU-Net-DSD: Depthwise Separable nnU-Net with DSD training strategy.

9. Results (2) – Qualitative Comparison
- The masks of our method are close to the ground truth labels, which further shows the feasibility of the
proposed method for efficient liver and lesion segmentation

10. Conclusions
- In this work, we design a novel end-to-end deep learning framework from both perspectives of data
and network for liver and tumor segmentation.
- Extensive experiments show that the proposed approach can obtain accurate segmentation results, as
well as speed up the training and inference process with only 7 M parameters.
- Ablation studies demonstrate the effectiveness of different components in our framework.

11. Thanks
Zhao_Ziyuan@i2r.a-star.edu.sg