Vessel segmentation from volumetric images: a multi-scale double-pathway network with class-balanced loss at the...

Vessel segmentation from volumetric medical images is becoming an essential pre-step in aiding the diagnosis, guiding the therapy, and patient management for vascular-related diseases. Deep learning-based methods have drawn many attentions, but most of them did not fully utilize the multi-scale spatial information of vessels. To address this shortcoming, we propose a multi-scale network similar to the well-known multi-scale DeepMedic. It also includes a double-pathway architecture and a class-balanced loss at the voxel level (MDNet-Vb) to achieve both the computation efficiency and segmentation accuracy.

The proposed network consists two parallel pathways to learn the multi-scale vessel morphology. Specifically, the pathway with a normal resolution uses three-dimensional (3D) U-Net fed with small inputs to learn the local details with relatively small storage and time consumption. The pathway with a low-resolution employs 3D fully convolutional network (FCN) fed with downsampled large inputs to learn the overall spatial relationships between vessels and adjacent tissues, and the morphological information of large vessels. To cope with the class-imbalanced issue in vessel segmentation, we propose a class-balanced loss at the voxel level with uniform sampling strategy. The class-balanced loss at the voxel level re-balances the loss function with a coefficient that is inversely proportional to the normalized effective number at the voxel level of each class. The uniform sampling strategy extracts training data by sampling uniformly from two classes in every epoch.

Our MDNet-Vb outperforms several state-of-the-art methods including ResNet, DenseNet, 3D U-Net, V-Net, and DeepMedic with the highest dice coefficients of 72.91% and 69.32% on cardiac computed tomography angiography (CTA) dataset and cerebral magnetic resonance angiography (MRA) dataset, respectively. Among four different double-pathway networks, our network (3D U-Net+3D FCN) not only has the fewest training parameters and shortest training time, but also gets competitive dice coefficients on both the CTA and MRA datasets. Compared with classical losses, our class-balanced focal loss (FL-Vb) and dice coefficient loss at the voxel level (Dsc-Vb) alleviates class imbalanced issue by improving both the sensitivity and dice coefficient on the CTA and MRA datasets. Moreover, simultaneously training on two datasets shows that our method has the highest dice coefficient of 73.06% and 65.40% on CTA and MRA datasets, respectively, outperforming the commonly used methods, such as U-Net and DeepMedic, which demonstrates the generalization potential of our network for segmenting different blood vessels.

Our MDNet-Vb method demonstrates its superiority over other state-of-the-art methods, on both cardiac CTA and cerebral MRA datasets. For the network architecture, the MDNet-Vb combined the 3D U-Net and 3D FCN, which dramatically reduces the network parameters yet maintains the segmentation accuracy. The class-balanced loss at the voxel level further improves accuracy by properly alleviating the class-imbalanced issue between different classes. In summary, MDNet-Vb is promising for vessel segmentation from various volumetric medical images.