RD4US: Unsupervised anomaly detection for deep vein thrombosis via cross-frame reverse distillation

Deep Vein Thrombosis (DVT) is a critical vascular condition conventionally diagnosed via Compression Ultrasonography (CUS), where the vein's compressibility serves as the primary biomarker. However, automating CUS analysis is hindered by the extreme scarcity of pathological samples and the prohibitive cost of expert annotation. To overcome these barriers, this paper proposes a novel unsupervised anomaly detection framework that reformulates the diagnostic task: rather than learning the variable appearance of thrombosis, the model learns the physiological consistency of healthy venous collapse. We introduce RD4US, a novel cross-frame reverse distillation framework tailored for this task. It leverages a pressure-sensor-equipped probe to acquire diagnostic triplets for each examination. Technically, our framework diverges from standard reconstruction methods by implementing a contrastive cross-reconstruction mechanism. This enforces the model to capture the semantic and morphological consistency of the vein, while a composite loss function actively suppresses confounding background artifacts inherent in ultrasound imaging. This design allows the detection of compressible veins in the feature space without requiring pixel-level segmentation masks. Comprehensive experiments on three clinical datasets demonstrate that our method achieves superior performance across diverse anatomical sites and clinical settings, consistently outperforming existing image-reconstruction and feature-reconstruction baselines with a classification accuracy exceeding 96% on the primary dataset. Notably, a reader study confirms that our model significantly elevates junior radiologists’ diagnostic accuracy to expert-level benchmarks, validating its clinical decision-making support capability. Furthermore, the proposed method generates precise anomaly localization heatmaps, providing interpretable visual support for clinical decision-making. Code is available at: https://github.com/ArkoDR/RD4US.