ECAU-Net: Efficient channel attention U-Net for fetal ultrasound cerebellum segmentation

• For the segmentation task of fetal US cerebellum images, a dataset termed as JSUAH-Cerebellum is constructed. • Unet is used as the backbone and combined with the ECA module to form a novel deep network ECA-Unet for US cerebellum segmentation. • By embedding 27 extra parameters, the ECA-Unet could fully utilize the low-level features and accurately segment the fetal cerebellum as well. • We reproduce 12 state-of-the-art algorithms and conduct comprehensive experiments. Ultrasound (US) examination of the fetal central nervous system is one of the significant tasks in the mid-term pregnancy inspection, in which the fetal cerebellum, as the key structure, attracts more attention from physicians. Therefore, accurate identification and segmentation of the cerebellum are particularly crucial for clinical diagnosis. In recent years, the development of deep learning has greatly improved the accuracy of computer automatic segmentation. This paper proposes an efficient channel attention U-Net algorithm for cerebellum segmentation (termed as ECAU-Net). The method employs the U-Net, which applies the encoder to extract feature representations and the decoder to locate segmentation results, as the backbone for segmentation. Combined with ECA modules, the one-dimensional convolutional layers with shared parameters are applied to replace the full connection layers in the conventional channel attention modules, which greatly reduces the number of model parameters without affecting performance. Moreover, we establish a US database, termed as JSUAH-Cerebellum, of fetal cerebellum images in the second trimester. During the training process, Gaussian blur and random flipping are utilized to boost the scale of the dataset. Comprehensive experiments were carried out with the Dice loss function on JSUAH-Cerebellum. In addition, we compared the proposed method with 12 other typical methods. The experimental results show that the mean Jaccard Score (JS) and Dice Similarity Coefficient (DSC) of the proposed algorithm are 86.01% and 91.35%, respectively, which are 2.86% and 3.73% higher than the baseline. Furthermore, for the worst-quality samples, the JS and DSC reach 0.55 and 0.68, respectively.