MHL-Net: A Multistage Hierarchical Learning Network for Head and Neck Multiorgan Segmentation

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• Accurate segmentation of head and neck organs at risk is critical for effective radiotherapy planning.
• Current segmentation methods struggle with feature extraction, information utilization, and handling organs of varying sizes.

Purpose of the Study:

• To develop an advanced deep learning network for improved segmentation of head and neck organs at risk.
• To address limitations in existing methods regarding feature mining and multiscale organ segmentation.

Main Methods:

• A multistage hierarchical learning network with multilevel subnetworks for primary, localization, and fine segmentation of large and small organs.
• Integration of anatomical prior probability maps and a boundary contour attention mechanism to handle complex organ shapes.
• A multidimensional combination attention mechanism to analyze 3D medical image data across axial, coronal, and sagittal planes.

Main Results:

• The proposed network demonstrated competitive performance against state-of-the-art methods.
• Significant improvements in segmentation accuracy were achieved for multiscale organs in the head and neck region.
• The method effectively utilized multidimensional features, anatomical priors, and attention mechanisms.

Conclusions:

• The multistage hierarchical learning network offers a robust solution for accurate head and neck organ segmentation in radiotherapy.
• The integration of anatomical priors and advanced attention mechanisms enhances segmentation of complex and multiscale structures.
• This approach holds promise for improving radiotherapy planning and delivery.