Cross-Domain Synthesis of Medical Images Using Efficient Location-Sensitive Deep Network
Cross-modality image synthesis has recently gained significant interest in the medical imaging community. In this paper, we propose a novel architecture called location-sensitive deep network (LSDN) for synthesizing images across domains. Our network inte
- PDF / 259,140 Bytes
- 8 Pages / 439.363 x 666.131 pts Page_size
- 57 Downloads / 182 Views
Abstract. Cross-modality image synthesis has recently gained significant interest in the medical imaging community. In this paper, we propose a novel architecture called location-sensitive deep network (LSDN) for synthesizing images across domains. Our network integrates intensity feature from image voxels and spatial information in a principled manner. Specifically, LSDN models hidden nodes as products of features and spatial responses. We then propose a novel method, called ShrinkConnect, for reducing the computations of LSDN without sacrificing synthesis accuracy. ShrinkConnect enforces simultaneous sparsity to find a compact set of functions that accurately approximates the responses of all hidden nodes. Experimental results demonstrate that LSDN+ShrinkConnect outperforms the state of the art in cross-domain synthesis of MRI brain scans by a significant margin. Our approach is also computationally efficient, e.g. 26× faster than other sparse representation based methods.
1
Introduction
Recently, cross-modality synthesis has gained significant interest in the medical imaging community. Existing approaches do not have a systematic way to incorporate the spatial information which is important for accurate synthesis. As an illustration, we plot the intensity correspondences of registered MRI-T1 and MRI-T2 of the same subject in Fig. 1a. We can notice that the intensity transformation is not only non-linear but also far from unique, i.e. there are multiple feasible intensity values in MRI-T2 domain for one intensity value in MRI-T1 domain. The non-uniqueness comes from a well-known fact that intensity values depend on the regions in which voxels reside. By restricting to a local neighborhood of, say 10 × 10 × 10 voxels, the intensity transformation is much simpler as shown in Fig. 1b. In particular, it could be reasonably well described as a union of two linear subspaces represented by the two red lines. That is to say, the spatial information helps simplify the relations between modalites which in turn could enable more accurate prediction. In this paper, we propose a novel architecture called location-sensitive deep network (LSDN) to integrate image intensity features and spatial information in a principled manner. Our network models the responses of hidden nodes as the product of feature responses and spatial responses. In LSDN formulation, spatial information is used as soft constraints whose parameters are learned. © Springer International Publishing Switzerland 2015 N. Navab et al. (Eds.): MICCAI 2015, Part I, LNCS 9349, pp. 677–684, 2015. DOI: 10.1007/978-3-319-24553-9_83
678
H. Van Nguyen, K. Zhou, and R. Vemulapalli 2000
2000
(b)
(a) T2 intensities
T2 Intensities
1500
1500
1000
1000
500
500
0
0
500
1000
1500
2000
2500
0
500
1000
1500
2000
2500
T1 Intensities
T1 Intensities
Fig. 1. a) 2D histogram of intensity correspondences between T1 and T2 scans over an entire image. Brighter color indicates higher density regions. b) Intensity correspondences of a restricted region of 10 × 10 × 10
Data Loading...
