Q. Dou, H. Chen, L. Yu, L. Zhao, J. Qin, D. Wang, V. C. T. Mok, L. Shi, P.-A. Heng. Automatic Detection of Cerebral Microbleeds From MR Images via 3D Convolutional Neural Networks. IEEE Trans. Med. Imaging, 2016.
Dou et al. propose an efficient approach to cerebral microbleeds detection in 3D MR volumes/images using 3D convolutional networks. Cerebral microbleeds are small areas of blood products within normal brain tissues. They have been identified as important symptom for diagnosing cerebrovascular diseases. However, manual detection is error prone, motivating the use of computer vision based systems for detection. An example can be found in Figure 1.
Figure 1 (click to enlarge): Illustration of an MR scan and a cerebral microbleed (in yellow) and a mimic (in red).
The proposed approach is a two-stage system consisting of a fully convolutional detection proposal system, and a discriminator to reduce false positives. In both cases, convolutional neural networks are generalized to 3D data in the straight-forward manner (by using 3D kernels for convolutions). The used architectures of both models are summarized in Figure 2, where $M$ denotes a max pooling layer, $C$ a convolutional layer and $FC$ a fully connected layer. For the fully convolutional network, the fully connected layers are reinterpreted as convolutional layers. Thus, the network can be trained on positive/negative crops of the 3D data and during testing be applied to full 3D volumes to produce probability volumes. After non-maximum suppression and thresholding, the probability volume is used to extract detection proposals. The discriminator is a general 3D convolutional neural network trained on crops including false positives obtained from the fully convolutional network. The full system and training procedure are illustrated in Figure 3.
Figure 2 (click to enlarge): Network architectures for the fully convolutional detection proposal network (left) and the discriminator network (right).
Figure 3 (click to enlarge): Illustration of training and testing of the full system.
On a newly created dataset of MR images, they show promising results (as far as I can judge) and demonstrate superiority regarding hand-crafted features, combined with random forests or SVMs. They also evaluate the detection proposal network separately to show that the number of false positives and false negatives is reduced compared to other approaches.
What is your opinion on this article? Let me know your thoughts on Twitter @davidstutz92 or LinkedIn in/davidstutz92.
Dou et al. propose an efficient approach to cerebral microbleeds detection in 3D MR volumes/images using 3D convolutional networks. Cerebral microbleeds are small areas of blood products within normal brain tissues. They have been identified as important symptom for diagnosing cerebrovascular diseases. However, manual detection is error prone, motivating the use of computer vision based systems for detection. An example can be found in Figure 1.
Figure 1 (click to enlarge): Illustration of an MR scan and a cerebral microbleed (in yellow) and a mimic (in red).
The proposed approach is a two-stage system consisting of a fully convolutional detection proposal system, and a discriminator to reduce false positives. In both cases, convolutional neural networks are generalized to 3D data in the straight-forward manner (by using 3D kernels for convolutions). The used architectures of both models are summarized in Figure 2, where $M$ denotes a max pooling layer, $C$ a convolutional layer and $FC$ a fully connected layer. For the fully convolutional network, the fully connected layers are reinterpreted as convolutional layers. Thus, the network can be trained on positive/negative crops of the 3D data and during testing be applied to full 3D volumes to produce probability volumes. After non-maximum suppression and thresholding, the probability volume is used to extract detection proposals. The discriminator is a general 3D convolutional neural network trained on crops including false positives obtained from the fully convolutional network. The full system and training procedure are illustrated in Figure 3.
Figure 2 (click to enlarge): Network architectures for the fully convolutional detection proposal network (left) and the discriminator network (right).
Figure 3 (click to enlarge): Illustration of training and testing of the full system.
On a newly created dataset of MR images, they show promising results (as far as I can judge) and demonstrate superiority regarding hand-crafted features, combined with random forests or SVMs. They also evaluate the detection proposal network separately to show that the number of false positives and false negatives is reduced compared to other approaches.