Segmentation of MR images via discriminative dictionary learning and sparse coding: Application to hippocampus labeling

TLDR: A novel method for the automatic segmentation of brain MRI images by using discriminative dictionary learning and sparse coding techniques, which can learn dictionaries offline and perform segmentations online, enabling a significant speed-up in the segmentation stage.

About: This article is published in NeuroImage.The article was published on 2013-08-01 and is currently open access. It has received 213 citations till now. The article focuses on the topics: Scale-space segmentation & Image segmentation.

Figures

Figure 8: The performance of Fixed Discriminative Dictionary Learning on 80 ICBM subjects. The results were obtained by using a patch size of 5× 5× 5 voxels and a search volume of 7 × 7 × 7 voxels. The experiment was repeated 10 times. The average median Dice coefficient is 0.887.

Figure 9: Effect of the number of training atlases on the performance of Fixed Discriminative Dictionary Learning. The results were obtained by using a patch size of 5× 5× 5 voxels and a search volume of 7× 7× 7 voxels. The median Dice coefficient is 0.864 when using 30 subjects for offline training and the remaining 172 subjects for online testing.

Figure 4: Effect of sampling step size on segmentation accuracy. The results were obtained by using a patch size of 5× 5× 5 voxels and a search volume of 7× 7× 7 voxels, extracted from the 10 most similar atlases.

Figure 3: Effect of Dictionary size on segmentation accuracy. The results were obtained by using a patch size of 5×5×5 voxels and a search volume of 7×7×7 voxels, extracted from the 10 most similar atlases. The sampling step size was set to 3 for learning dictionaries.

Table 2: Median Dice overlaps for 80 ICBM subjects. The results of F-DDLS* were obtained by randomly selecting 40 atlases for training and using the remaining 40 subjects for evluation. The experiment was repeated 10 times. The other results were obtained by using the most similar 20 atlases in a leave-one-out procedure. The Patch-based and SRC methods were implemented by using a patch size of 7×7×7 voxels and a search volume of 9× 9× 9 voxels. The DDLS and F-DDLS methods were implemented by using a sampling step size of 3, a patch size of 5 × 5 × 5 voxels and a search volume of 7 × 7 × 7 voxels. The difference between Patch-based and DDLS is statistically significant with p<0.001 on Students two-tailed paired t-test.

Table 3: Median Dice overlaps for 202 ADNI subjects. The results of F-DDLS* were obtained by randomly selecting 30 atlases for training and using the remaining 172 subjects for evluation. The other results were obtained by using the 10 most similar atlases in a leave-one-out procedure. The Patch-based and SRC methods were implemented by using a patch size of 7 × 7 × 7 voxels and a search volume of 9 × 9 × 9 voxels. The DDLS and F-DDLS methods were implemented by using a sampling step size of 3, a patch size of 5×5×5 voxels and a search volume of 7×7×7 voxels. The difference between Patch-based and DDLS is statistically significant with p<0.001 on Students two-tailed paired t-test.