Multi-Slice Dense-Sparse Learning for Efficient Liver and Tumor Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Pathology

Background:

• Accurate liver and tumor segmentation is crucial for cancer treatment planning and monitoring.
• Deep convolutional neural networks (DCNNs) show promise in medical image segmentation.
• Existing 2D DCNNs miss inter-slice data, while 3D DCNNs are resource-intensive.

Purpose of the Study:

• To develop an efficient and accurate method for automatic liver and tumor segmentation from CT scans.
• To address the limitations of 2D and 3D DCNNs in medical image segmentation.
• To improve DCNN performance by leveraging both dense and sparse slice information.

Main Methods:

• A novel dense-sparse training strategy was proposed to regularize DCNNs using adjacent and sparse slices.
• A lightweight 2.5D nnU-Net architecture was designed incorporating depthwise separable convolutions for enhanced efficiency.
• The proposed methods were evaluated using the Liver Tumor Segmentation (LiTS) dataset.

Main Results:

• The proposed dense-sparse training flow significantly improved DCNN model performance.
• The 2.5D light-weight nnU-Net demonstrated superior efficiency compared to traditional methods.
• Experiments on the LiTS dataset confirmed the effectiveness and superiority of the developed approach.

Conclusions:

• The novel training flow and network architecture offer an effective solution for liver and tumor segmentation.
• The method achieves high accuracy with reduced computational complexity, facilitating clinical implementation.
• This work advances automated medical image analysis for improved patient care and disease management.