DHT-Net: Dynamic Hierarchical Transformer Network for Liver and Tumor Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Convolutional Neural Networks (CNNs) struggle with long-range dependencies for complex tumor feature extraction in liver tumor segmentation.
• Existing Transformer-based 3D networks often focus on either local or global information with fixed weights, limiting adaptability to varied tumor characteristics.

Purpose of the Study:

• To develop a Dynamic Hierarchical Transformer Network (DHT-Net) for accurate automatic liver tumor segmentation.
• To enhance the extraction of complex tumor features, considering variations in size, location, and morphology.

Main Methods:

• Proposed DHT-Net integrates a Dynamic Hierarchical Transformer (DHTrans) and an Edge Aggregation Block (EAB).
• DHTrans utilizes Dynamic Adaptive Convolution with hierarchical operations and varying receptive fields to learn diverse tumor features and aggregate global/local texture information.
• EAB extracts fine-grained edge features for precise boundary delineation.

Main Results:

• DHT-Net demonstrated superior performance in liver and tumor segmentation on the LiTS and 3DIRCADb datasets.
• The method outperformed several state-of-the-art 2D, 3D, and 2.5D hybrid models in segmentation accuracy.

Conclusions:

• DHT-Net effectively addresses limitations of previous methods by dynamically learning complex tumor features.
• The proposed network offers improved accuracy for automatic liver tumor segmentation, aiding radiologists in clinical diagnosis.