An Efficient Multi-Scale Fusion Network for 3D Organs at Risk (OARs) Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiation Oncology

Background:

• Accurate segmentation of organs-at-risk (OARs) is crucial for radiation therapy planning.
• Deep learning excels in 2D medical image segmentation but faces challenges in 3D due to computational demands and data requirements.
• Existing 3D multi-scale fusion architectures may underperform.

Purpose of the Study:

• To develop an efficient and effective 3D segmentation framework for organs-at-risk.
• To address the limitations of current 3D multi-scale fusion architectures in medical image segmentation.

Main Methods:

• Proposed OARFocalFuseNet, a novel framework for OAR segmentation.
• Implemented multi-scale feature fusion with focal modulation for global-local context capture.
• Enriched resolution streams with features from different scales and aggregated multi-scale information.

Main Results:

• OARFocalFuseNet demonstrated superior performance compared to state-of-the-art methods on OpenKBP and Synapse datasets.
• Achieved a Dice coefficient of 0.7995 and Hausdorff distance of 5.1435 on OpenKBP datasets.
• Attained a Dice coefficient of 0.8137 on the Synapse multi-organ segmentation dataset.

Conclusions:

• OARFocalFuseNet effectively fuses multi-scale features and utilizes focal modulation for enhanced 3D OAR segmentation.
• The proposed method offers a promising solution for optimizing radiation therapy planning through improved segmentation accuracy.