Anatomy-aware fusion attention network for high-precision prostate MRI segmentation.

Background

Accurate prostate segmentation in Magnetic Resonance Imaging (MRI) is crucial for clinical diagnosis and treatment. However, it remains a challenging task due to low soft-tissue contrast and structural ambiguity in the gland's appearance. While recent deep learning methods have focused on refining network structures, they often neglect the inherent anatomical consistency of the prostate.

Purpose

To address this limitation, a novel segmentation framework integrating an Anatomy-Aware Fusion Attention (AAFA) module is proposed.

Methods

By leveraging anatomical templates, our approach establishes a multilevel feature cross-attention mechanism that enhances global contextual modeling of prostate regions. Additionally, we introduce a Phased Learning strategy that progressively trains the model to mitigate the adverse effects of invalid or noisy samples commonly found in clinical MRI data.

Results

Extensive ablation studies on the PROMISE12 dataset validate the contribution of each component to overall performance. Comparison experiments on both PROMISE12 and MSD Prostate datasets show that our method consistently outperforms existing approaches across key metrics, such as Dice similarity coefficient (DSC), Intersection over Union (IoU), Precision and 95% Hausdorff distance (HD95).

Conclusions

These results confirm the robustness and strong generalization capability of the proposed framework in challenging clinical segmentation tasks.