A novel thyroid nodule ultrasound image segmentation method based on UNet++

In response to the current challenges in accurately segmenting thyroid nodules in ultrasound images, this study introduces an enhanced thyroid nodule segmentation approach for ultrasound images, developed through architectural modifications to the UNet++ framework. The method enhances the original model by incorporating an efficient channel attention module, which adaptively adjusts the feature weights of channels with different receptive fields, allowing channels with useful information to dominate. Additionally, the Dense Atrous Spatial Pyramid Pooling module is introduced to help the network utilize more pixels, achieve a larger receptive field, and avoid the inefficiency or failure of atrous convolutions. Quantitative evaluation demonstrates that our approach attains a Dice coefficient of 0.8904 on thyroid ultrasound segmentation tasks, demonstrating superior segmentation performance compared to traditional methods.