PUFormer: Prior Knowledge-Guided and Uncertainty-Driven Multi-Scale Fusion Network for Head and Neck OAR Segmentation

Abstract

Computer-aided diagnosis plays a pivotal role in clinical medicine. In recent years, models based on convolutional neural networks (CNNs) and Transformers have achieved remarkable performance in medical image segmentation. However, existing methods still face significant challenges in multi-organ segmentation tasks, particularly when dealing with anatomically complex structures and targets exhibiting large shape variations. To address these limitations, we propose PUFormer, a novel CNN-Transformer hybrid architecture. The proposed model consists of three key components: (1) a prior knowledge-guided learning strategy, implemented through Prior Shunted Transformer Blocks, each of which incorporates a Prior Shunted Self-Attention (PSSA) mechanism as its core operation, leveraging anatomical prior knowledge to guide the model in learning consistent spatial relationships among organs; (2) a feature aggregation enhancement module, designed to extract effective multi-scale features from aggregated multi-level representations; and (3) an uncertainty-driven multi-scale gated fusion module, which employs prediction entropy-based uncertainty to dynamically allocate fusion weights, thereby improving the effectiveness of multi-scale feature integration. We conducted extensive experiments on three public datasets: HaN-Seg, SegRap2023, and ACDC. The results demonstrate that PUFormer achieves competitive performance compared to state-of-the-art methods. Specifically, on the HaN-Seg dataset, it attains an average Dice coefficient of 77.04% and HD95 of 3.56 mm. On SegRap2023, it reaches an average Dice of 87.02% and HD95 of 3.04 mm. On the ACDC dataset, the model achieves an average Dice coefficient of 92.65% and HD95 of 3.22 mm. These findings validate the effectiveness of PUFormer in head and neck organs-at-risk (OAR) segmentation, and further demonstrate its applicability across different medical imaging modalities.