Automatic Multi-Organ Segmentation on Abdominal CT With Dense V-Networks
read more
Citations
CE-Net: Context Encoder Network for 2D Medical Image Segmentation
Deep Learning Techniques for Medical Image Segmentation: Achievements and Challenges
CE-Net: Context Encoder Network for 2D Medical Image Segmentation
NiftyNet: a deep-learning platform for medical imaging
UNETR: Transformers for 3D Medical Image Segmentation
References
Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift
Fully convolutional networks for semantic segmentation
Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift
Rectified Linear Units Improve Restricted Boltzmann Machines
A survey on deep learning in medical image analysis
Related Papers (5)
Frequently Asked Questions (17)
Q2. What are the future works in "Automatic multi-organ segmentation on abdominal ct with dense v-networks" ?
The evaluation metrics measure segmentation fidelity with the manual reference, and not the clinical utility of the resulting segmentations for aiding endoscopic navigation ; future work will evaluate whether the proposed algorithm is accurate enough to provide a 3D patientspecific anatomical model to aid endoscopic navigation. The use of dilated convolutions was not necessary, suggesting that global high-resolution non-linear features are not critical for abdominal CT organ segmentation. The use of an explicit spatial prior was also not necessary, suggesting that convolutional neural networks are implicitly encoding spatial priors, despite their purported translational invariance. The automatically generated segmentations of abdominal anatomy have the potential to support image-guided navigation in pancreatobiliary endoscopy procedures.
Q3. What features were introduced to reduce memory costs without affecting performance?
Two features were introduced to reduce memory costs without affecting performance: batch-wise spatial dropout and Monte Carlo inference.
Q4. What is the way to achieve memory efficiency in dense blocks?
Memory-efficient dense blocks [44], where a careful implementation of feature concatenation avoids storing multiple copies of feature maps, can achieve O(m) memory usage.
Q5. How can the authors reduce the memory usage of the sample?
MonteCarlo inference [41] can be used (increasing the computation cost but lowering the memory usage) by inferring multiple segmentation samples using dropout, and combining them.
Q6. What are the common uses of FCNs?
FCNs have recently been applied to segmentation of volumetric images in medical image analysis [18], [19], [24]–[26] where such images are common.
Q7. What is the common problem with the Dice coefficient?
The relative weighting of the losses for different organs (with high volume imbalance) can have unpredictable effects on convergence and final errors; using the Dice coefficient is common but remains poorly characterized.
Q8. What can be the impact of a cropping protocol on segmentation accuracy?
Altering the cropping protocol for the test data sufficiently (i.e. beyond the variability generated by data augmentation) can impact segmentation accuracy.
Q9. What is the way to limit the memory usage of large volumetric images?
One strategy to constrain the memory usage is to process smaller images: small patches of a larger image or lower resolution images.
Q10. What are the common multi-organ segmentation methodologies?
1) Common multi-organ segmentation methodologies: Statistical models [5], [6] involve co-registering images in a training data set to estimate anatomical correspondences, constructing a statistical model of the distribution of shapes [22] and/or appearances [23] of corresponding anatomy in the training data, and fitting the resulting model to new images to generate segmentations.
Q11. What are the challenges of segmentation of volumetric images?
Segmentation of volumetric images face particular challenges, mainly due to the need to process large volumetric images under memory constraints.
Q12. What are the limitations of the proposed deep learning method?
In conclusion, the proposed deep-learning-based DenseVNet can segment the pancreas, esophagus, stomach, liver, spleen, gallbladder, left kidney and duodenum more accurately than previous methods using deep learning or multi-atlas label fusion.
Q13. What is the name of the algorithm used to evaluate Monte Carlo inference?
To evaluate Monte Carlo inference, the trained DenseVNet was used, but inference was performed with no dropout, using all features; these results are abbreviated as Deterministic.
Q14. What is the probability distribution of keeping k out of n channels?
In spatial dropout, the probability distribution of keeping k out of n channels is a binomial distribution p(K = k) = ( n k ) pk(1 − p)n−k; although the expected value E[K = k] = pn, the maximum value (corresponding to the maximum memory usage) is n.
Q15. What is the way to navigate an endoscope?
Bottom: Segmentations overlaid on CT.targets) and the gastrointestinal tract (where the endoscope is navigated) should be prioritized over navigational landmarks as an endoscope can be oriented without precise boundaries.
Q16. What are the limitations of the deep learning methods?
Although these times would not be a limiting factor in a clinical workflow for fully-automated segmentation, the deep learning methods are fast enough to use for more accurate semi-automatic segmentations.
Q17. What limitations did the authors have to overcome in the development of the algorithm?
Authors were not blinded to the manual segmentations during algorithm development; although the cross-validation was only run after algorithm development was complete, design decisions may have been influenced by data observations.