Medical image segmentation plays a vital role in computer-aided diagnosis procedures. Recently, U-Net is widely used in medical image segmentation. Many variants of U-Net have been proposed, which attempt to improve the network performance while keeping the U-shaped structure unchanged. However, this U-shaped structure is not necessarily optimal. In this article, the effects of different parts of the U-Net on the segmentation ability are experimentally analyzed. Then a more efficient architecture, Half-UNet, is proposed. The proposed architecture is essentially an encoder-decoder network based on the U-Net structure, in which both the encoder and decoder are simplified. The re-designed architecture takes advantage of the unification of channel numbers, full-scale feature fusion, and Ghost modules. We compared Half-UNet with U-Net and its variants across multiple medical image segmentation tasks: mammography segmentation, lung nodule segmentation in the CT images, and left ventricular MRI image segmentation. Experiments demonstrate that Half-UNet has similar segmentation accuracy compared U-Net and its variants, while the parameters and floating-point operations are reduced by 98.6 and 81.8%, respectively, compared with U-Net.

BackgroundImplementation of deep learning systems (DLSs) for analysis of barium esophagram, a cost-effective diagnostic test for esophageal cancer detection, is expected to reduce the burden to radiologists while ensuring the accuracy of diagnosis.ObjectiveTo develop an automated DLS to detect esophageal cancer on barium esophagram.MethodsThis was a retrospective study using deep learning for esophageal cancer detection. A two-stage DLS (including a Selection network and a Classification network) was developed. Five datasets based on barium esophagram were used for stepwise training, validation, and testing of the DLS. Datasets 1 and 2 were used to respectively train and test the Selection network, while Datasets 3, 4, and 5 were respectively used to train, validate, and test the Classification network. Finally, a positioning box with a probability value was outputted by the DLS. A region of interest delineated by experienced radiologists was selected as the ground truth to evaluate the detection and classification efficiency of the DLS. Standard machine learning metrics (accuracy, recall, precision, sensitivity, and specificity) were calculated. A comparison with the conventional visual inspection approach was also conducted.ResultsThe accuracy, sensitivity, and specificity of our DLS in detecting esophageal cancer were 90.3%, 92.5%, and 88.7%, respectively. With the aid of DLS, the radiologists’ interpretation time was significantly shortened (Reader1, 45.7 s vs. 72.2 s without DLS aid; Reader2, 54.1 s vs. 108.7 s without DLS aid). Respective diagnostic efficiencies for Reader1 with and without DLS aid were 96.8% vs. 89.3% for accuracy, 97.5% vs. 87.5% for sensitivity, 96.2% vs. 90.6% for specificity, and 0.969 vs. 0.890 for AUC. Respective diagnostic efficiencies for Reader2 with and without DLS aid were 95.7% vs. 88.2% for accuracy, 92.5% vs. 77.5% for sensitivity, 98.1% vs. 96.2% for specificity, and 0.953 vs. 0.869 for AUC. Of note, the positioning boxes outputted by the DLS almost overlapped with those manually labeled by the radiologists on Dataset 5.ConclusionsThe proposed two-stage DLS for detecting esophageal cancer on barium esophagram could effectively shorten the interpretation time with an excellent diagnostic performance. It may well assist radiologists in clinical practice to reduce their burden.