TFNet: Transformer Fusion Network for Ultrasound Image Segmentation

Automatic lesion segmentation in ultrasound helps diagnose diseases. Segmenting lesion regions accurately from ultrasound images is a challenging task due to the difference in the scale of the lesion and the uneven intensity distribution in the lesion area. Recently, Convolutional Neural Networks have achieved tremendous success on medical image segmentation tasks. However, due to the inherent locality of convolution operations, it is limited in modeling long-range dependency. In this paper, we study the more challenging problem on capturing long-range dependencies and multi-scale targets without losing detailed information. We propose a Transformer-based feature fusion network (TFNet), which fuses long-range dependency of multi-scale CNN features via Transformer to effectively solve the above challenges. In order to make up for the defect of Transformer in channel modeling, will be improved by joining the channel attention mechanism. In addition, a loss function is designed to modify the prediction map by computing the variance between the prediction results of the auxiliary classifier and the main classifier. We have conducted experiments on three data sets, and the results show that our proposed method achieves superior performances against various competing methods on ultrasound image segmentation.