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.
1 March 2024 in IEEE Journal of Biomedical and Health Informatics
Mammography is an effective method for diagnosing breast diseases, and computer-aided detection (CAD) systems play an important role in the detection of breast masses. However, low contrast and the interference of surrounding tissues make the detection of masses challenging. In this paper, an efficient RetinaNet network named ERetinaNet is proposed to improve the accuracy and inference speed of mammographic breast mass detection. Efficient modules are designed and introduced into the network to facilitate the extraction of comprehensive features, while the structure of the network is simplified to improve the inference speed. A Faster RepVGG (FRepVGG) architecture is first proposed as the backbone network that utilizes three effective strategies: (1) The multi-branch structure used during training enhances learning, and it is equivalently converted to a singlepath structure during inference by re-parameterization technique to accelerate the detection speed. (2) The Extraction operation is proposed to condense the features of intermediate layers. (3) An effective Multi-spectral Channel Attention (eMCA) module is added in the last layer of each stage, enabling the network to pay more attention to the target region. In addition, Vision Transformer (ViT) is added to ERetinaNet, which enables ERetinaNet to learn global semantic information. The detection head is simplified to make ERetinaNet more efficient. The experimental results show that compared with the original RetinaNet, ERetinaNet improves the mean Average Precision (mAP) from 79.16% to 85.01% and significantly shortens the inference time. Moreover, the detection accuracy of ERetinaNet outperforms other excellent object detection networks, such as Faster R-CNN, SSD, YOLOv3 and YOLOv7.
Background: Implementation 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.Objective: To develop an automated DLS to detect esophageal cancer on barium esophagram.Methods: This 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.Results: The 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.Conclusions: The 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.
26 March 2020 in International Journal of Imaging Systems and Technology
Cardiovascular diseases are reported as the leading cause of death around the world. Automatic segmentation of the left ventricle (LV) from magnetic resonance (MR) images is essential for an early diagnosis. An enhanced ResUnet is proposed in this paper to improve the performance of extracting LV endocardium and epicardium from MR images, improving the accuracy of the model by introducing a medium skip connection for the contracting path and a short skip connection for the residual unit. Also, a depth-wise separable convolution replaces the typical convolution operation to improve training efficiency. In the MICCAI 2009 LV segmentation challenge test dataset, the percentages of “good” contours, dice metric, and average perpendicular distance of endocardium (epicardium) are , respectively. Experimental results demonstrate that the proposed model obtains promising performance and outperforms state-of-the-art methods. By incorporating these various skip connections, the segmentation accuracy of the model is significantly improved, while the depth-wise separable convolution also improves the model efficiency.
The automatic segmentation of the left ventricle in magnetic resonance (MR) images is the basis of computer-aided diagnosis systems. To accurately extract the endocardium and epicardium of the left ventricle from MR images, a method based on a dilated dense convolutional network (DDCN) has been proposed in this paper. First, to reduce memory consumption, computing time and the class imbalance between the target and background, a clustering algorithm that combines the prior knowledge of the spatial relationship between the slices has been proposed to crop the region of interest (ROI). Then, the DDCN model with 8 dilated convolutional layers and dense connections, which is efficient with respect to its memory consumption and training time, has been proposed to delineate the endocardium and epicardium. To compare the DDCN model with other algorithms, 30 sequences of the MICCAI 2009 left ventricle segmentation challenge database are used to train the proposed model and the other 15 sequences are used for testing. The performance of the proposed method is evaluated by the percentage of "good" contours (PGC), average Dice metric (ADM) and average perpendicular distance (APD). Our results show that for the endocardial and epicardial contours, the PGCs are 99.49%±1.99% and 100%±0%, the APDs are 1.50 ± 0.34 mm and 1.31 ± 0.22 mm, and the ADMs are 0.93 ± 0.03 and 0.96 ± 0.01, respectively, which indicates that our method provides contours with great agreement with the ground truth. In addition, the comparison results show that our method exhibits outstanding performance and possesses promising potential to be used in computer-aided diagnosis systems for cardiovascular disease.