MAD-Net: Multi-attention dense network for functional bone marrow segmentation.

Area of Science:

• Medical Physics
• Artificial Intelligence in Medicine
• Oncology

Background:

• Radiotherapy for pelvic malignancies can damage functional bone marrow (FBM), leading to hematological toxicity (HT).
• Manual FBM segmentation for radiotherapy planning is inefficient and labor-intensive.
• Accurate FBM identification is crucial for minimizing pelvic HT.

Purpose of the Study:

• To develop an automated, efficient, and accurate method for FBM segmentation in pelvic radiotherapy.
• To introduce the Multi-Attention Dense Network (MAD-Net) for FBM segmentation.

Main Methods:

• Proposed MAD-Net incorporates dense convolution blocks for enhanced gradient flow and feature reuse.
• Introduced a novel slide-window attention module to capture long-range dependencies.
• Utilized a residual-dual attention module in the bottleneck layer for spatial detail aggregation and feature responsiveness.

Main Results:

• MAD-Net demonstrated superior performance compared to state-of-the-art models on a dataset of 3838 pelvic slices.
• Ablation studies validated the effectiveness of individual components within MAD-Net.
• Experiments on three additional datasets confirmed the generalizability of MAD-Net.

Conclusions:

• MAD-Net represents a significant advancement in automated FBM segmentation for radiotherapy planning.
• The proposed network offers a promising solution for reducing hematological toxicity in patients undergoing pelvic radiotherapy.
• MAD-Net's generalizability suggests its broad applicability in clinical settings.