Prior-Guided DETR for Ultrasound Nodule Detection

TL;DR

Ultrasound nodules exhibit irregular morphology, blurred boundaries, speckle noise, and multi-scale variations that challenge purely data-driven detectors. The authors introduce a prior-guided DETR framework that progressively injects domain knowledge through three components: SDFPR for geometric priors in deformable sampling, MSFFM for spatial-frequency structural priors, and DFI for dense, multi-level feature interaction with the decoder. Across internal Thyroid I/II data and external TN3K and BUSI datasets, the method achieves state-of-the-art performance, with pronounced gains on small and large nodules and strong cross-organ generalization. This work demonstrates the value of explicit priors in medical object detection, offering robustness and potential to streamline clinical workflows with real-time applicability on portable devices.

Abstract

Accurate detection of ultrasound nodules is essential for the early diagnosis and treatment of thyroid and breast cancers. However, this task remains challenging due to irregular nodule shapes, indistinct boundaries, substantial scale variations, and the presence of speckle noise that degrades structural visibility. To address these challenges, we propose a prior-guided DETR framework specifically designed for ultrasound nodule detection. Instead of relying on purely data-driven feature learning, the proposed framework progressively incorporates different prior knowledge at multiple stages of the network. First, a Spatially-adaptive Deformable FFN with Prior Regularization (SDFPR) is embedded into the CNN backbone to inject geometric priors into deformable sampling, stabilizing feature extraction for irregular and blurred nodules. Second, a Multi-scale Spatial-Frequency Feature Mixer (MSFFM) is designed to extract multi-scale structural priors, where spatial-domain processing emphasizes contour continuity and boundary cues, while frequency-domain modeling captures global morphology and suppresses speckle noise. Furthermore, a Dense Feature Interaction (DFI) mechanism propagates and exploits these prior-modulated features across all encoder layers, enabling the decoder to enhance query refinement under consistent geometric and structural guidance. Experiments conducted on two clinically collected thyroid ultrasound datasets (Thyroid I and Thyroid II) and two public benchmarks (TN3K and BUSI) for thyroid and breast nodules demonstrate that the proposed method achieves superior accuracy compared with 18 detection methods, particularly in detecting morphologically complex nodules.The source code is publicly available at https://github.com/wjj1wjj/Ultrasound-DETR.

Figures (8)

• Figure 1: Typical challenges in ultrasound nodule detection. Irregular morphology, blurred boundaries, and strong speckle noise jointly degrade structural visibility, while pronounced scale variation further complicates robust localization. These characteristics reveal a fundamental mismatch between ultrasound image formation physics and purely data-driven detection models, motivating the explicit incorporation of geometric and structural priors into the detection framework.
• Figure 2: Overview of the proposed prior-guided DETR. To address the mismatch between ultrasound physics and implicit feature learning, our framework progressively injects domain knowledge at three hierarchical stages: 1) SDFPR embeds Geometric Prior into the backbone to stabilize deformable sampling; 2) MSFFM extracts Structural Prior by synergizing spatial boundary cues with frequency-domain morphology; and 3) DFI propagates these prior-modulated features to the decoder via dense interaction. This unified paradigm ensures robust detection against irregular morphology, speckle noise and multi-scale variation.
• Figure 3: Structure of the Spatially-adaptive Deformable FFN with Prior Regularization (SDFPR). Conventional deformable convolutions often suffer from unstable offset regression. By embedding geometric priors (aspect ratio and width) learned from clinical data into deformable convolution, SDFPR regularizes offset learning and stabilizes geometric modeling for nodules with irregular shapes and blurred boundaries.
• Figure 4: Structure of the Multi-scale Spatial-Frequency Feature Mixer (MSFFM). MSFFM extracts structural priors by bridging two complementary domains: the Spatial Branch (top) aggregates local contour continuity via Perception-Aggregation Convolution, while the Frequency Branch (bottom) utilizes learnable spectral filtering to suppress speckle noise and highlight global morphology. An adaptive fusion strategy dynamically balances these local and global cues to generate robust representations across different nodule scales.
• Figure 5: (a) PAConv employs a dual-phase strategy to capture long-range spatial dependencies. (b) Frequency-Domain impact. Spatial features are transformed into the spectral domain, where learnable reweighting enhances morphology-related low-frequency components and suppresses speckle-dominated high-frequency noise. After inverse transformation, the reconstructed features exhibit cleaner backgrounds and more coherent nodule responses, reflecting a morphology prior.