Enterprise AI Research Analysis

Tracking spatial temporal details in ultrasound long video via wavelet analysis and memory bank

Authors: Chenxiao Zhang, Runshi Zhang, Junchen Wang

Medical ultrasound videos are crucial for disease diagnosis and surgical planning, but low contrast and noisy backgrounds lead to segmentation errors, especially for small objects and in long videos. To address this, we propose MWNet, a memory bank-based wavelet filtering and fusion network with an encoder-decoder structure. It leverages memory-based wavelet convolution for spatiotemporal feature extraction and cascaded wavelet compression for multi-scale frequency-domain fusion. A long short-term memory bank with cross-attention tracks objects across long videos. An HF-aware feature fusion module, using adaptive wavelet filters, enhances boundary-sensitive details. Extensive tests on four ultrasound video datasets show MWNet significantly improves segmentation metrics, particularly for small thyroid nodules, demonstrating superior effectiveness in challenging long ultrasound video scenarios.