Search

While modern deep learning methods have made significant progress in medical image segmentation, some challenges remain, including accurately capturing features at multiple scales, limited ability to detect critical regions, and susceptibility to noise and background interference. To address these challenges, a new neural network called HMSAM-UNet is introduced in this work. A novel module, the Hierarchical Multi-Scale Attention Module (HMSAM), was designed in HMSAM-UNet to improve the precision and accuracy of CT image segmentation significantly. Specifically, HMSAM integrates the Hierarchical Attention Mechanism and Inception Module via residual connections. The Hierarchical Attention Mechanism can highlight important regions by learning attention weights, dramatically enhancing the model’s ability to perceive critical areas for more accurate localization and segmentation of target structures in CT images. Meanwhile, incorporating the Inception module effectively strengthens the network’s capacity to capture multi-scale features, substantially improving the model’s ability to comprehend the structural characteristics of CT images. The results show that the average loss achieved by the proposed model has a 50.04% reduction compared to the original U-Net architecture. Furthermore, compared to other deep learning models such as FCN, DeepLabV3, PSPNet, Unet, UNet++, and SegNet, the model proposed in this work attains an average Dice coefficient of 98.72, and an average IoU score of 97.46 on the three datasets, both of which are the highest among all compared models.

Decent whole-heart segmentation from computed tomography (CT) can greatly contribute to the diagnosis and treatment of cardiovascular diseases. However, due to the difficulties such as blurred boundaries between neighbouring tissues and a large number of background voxels in medical images, automated whole-heart segmentation is still a challenging task. In this paper, we proposed three modified attention models, including simple negative example mining (SNEM), attention gate (AG) and U-CliqueNet (UCNet), to lead the deep learning network to focus on more salient information. These three attention modules were further implemented into a deeply-supervised 3D UNET separately and jointly, showing different degrees of improvement on the whole-heart segmentation task. Our experiments advised that SNEM was the most simple and effective attention mechanism for medical image processing among the three and the UCNet could reach the best performance. The combination of the attention mechanisms cannot always synergistically increase the accuracy, but joint models would have a positive influence in most cases. Finally, our network achieved a Dice score of 0.9112, which was a substantially higher performance than most of the state-of-the-art methods.

Obtaining precise whole-heart segmentation from computed tomography (CT) or other imaging techniques is prerequisite to clinically analyze the cardiac status, which plays an important role in the treatment of cardiovascular diseases. However, the whole-heart segmentation is still a challenging task due to the characteristic of medical images, such as far more background voxels than foreground voxels and the indistinct boundaries of adjacent tissues. In this paper, we first present a new deeply supervised 3D UNET which applies multi-depth fusion to the original network for a better extract context information. Then, we apply focal loss to the field of image segmentation and expand its application to multi-category tasks. Finally, the focal loss is incorporated into the Dice loss function (which can be used to solve category imbalance problem) to form a new loss function, which we call hybrid loss. We evaluate our new pipeline on the MICCAI 2017 whole-heart CT dataset, and it obtains a Dice score of 90.73%, which is better than most of the state-of-the-art methods.