A new ultrasound thyroid nodule segmentation model with multi-scale features and attention mechanisms is proposed to address the problem of inaccurate segmentation of thyroid nodules in ultrasound images.In the model,a multi receptive field channel selection module designed in the encoding phase is used to adaptively weight and combine features from different receptive fields with a focus on channels containing the target object.Additionally,an adaptive global context module is designed to extract global context features from multiple scales of the bottleneck layer to effectively encode high-level semantic information.Furthermore,a dual attention-guided module is used to enhance feature fusion between peer layers of the encoder and decoder for reducing information loss during upsampling.Experimental results on a publicly available ultrasound thyroid nodule dataset show that the proposed method outperforms other state-of-the-art networks and achieves more accurate segmentation of thyroid nodules,effectively improving segmentation performance.
关键词
深度学习/甲状腺结节/超声图像分割/多尺度特征提取/注意力机制
Key words
deep learning/thyroid nodule/ultrasound image segmentation/multi-scale feature extraction/attention mechanism
维普
万方数据