Doppler Image-Based Weakly-Supervised Vascular Ultrasound Segmentation with Transformers

Vascular segmentation in ultrasound (US) images faces labor-intensive labeling procedures and performance degradation due to unsatisfied image quality. Herein, we propose a Doppler image-based vascular segmentation method with a Transformer encoder-decoder structure. First, to extract global features of the US image, we utilize a Transformer encoder with a convolutional neural networks (CNN) patch embedding model. The extracted local features are further passed through a multi-level Transformer encoder to enhance the global dependencies. Finally, a multi-level CNN decoder is introduced to decode image features from coarse to fine. Doppler imaging is capable of blood visualization and indicating the positional and structural information of the vessel, which are used as the pseudo label. To improve the effectiveness of Doppler images, a CRF module and shape similarity loss function are introduced. The segmentation accuracy of two clinical dataset can achieve 78.8% and 81.9% in Dice with 63.5% and 53.7% accuracy in noisy labels.