End-to-End Adversarial Retinal Image Synthesis
read more
Citations
Generative adversarial network in medical imaging: A review.
Deep learning in medical imaging and radiation therapy.
Medical Image Synthesis for Data Augmentation and Anonymization Using Generative Adversarial Networks
Image Synthesis in Multi-Contrast MRI With Conditional Generative Adversarial Networks
Contrastive learning of global and local features for medical image segmentation with limited annotations
References
Generative Adversarial Nets
Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift
Auto-Encoding Variational Bayes
Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift
Image-to-Image Translation with Conditional Adversarial Networks
Related Papers (5)
Frequently Asked Questions (16)
Q2. What have the authors stated for future works in "End-to-end adversarial retinal image synthesis" ?
There are other limitations of the proposed approach that should be object of future research. Therefore, in the future, the introduction of clinical labels or annotations in the context of a large scale high-resolution data collection will be the first natural extension of their model, as a part of the more general goal of producing realistic and interesting synthetic images that can be employed to train models to solve more complex retinal image analysis tasks. Most of the above drawbacks can be attributed to the amount of available data and computational resource restrictions, and not to a limitation intrinsic to the proposed technique. In general, the availability of an additional set of training examples that can be efficiently generated on-demand could greatly impact the size and capacity of the models the retinal image analysis community train.
Q3. What is the main advantage of this joint training scheme?
The main advantage of this joint training scheme is that the discriminator D also provides with a better loss function for the adversarial autoencoder.
Q4. What is the performance of the vessel segmentation model when trained with real images?
The performance of the vessel segmentation model when trained with synthetic images is well above a baseline random model, and when allowed a fraction of false positives approximately greater than 0.35, the resulting system shows greater sensitivity than the same segmentation model trained with real images.
Q5. Why does the vessel network 2 show high plausibility?
The vessel network 2 also shows high plausibility, with the two main arcades displaying2Note that the synthetic vessel trees contain continuous values in [0,1], due to their model minimizing the cross-entropy loss.
Q6. What is the way to use the vessel networks?
These vessel networks can be considered as probability maps, and thresholded appropriately if a binary vessel network is needed for some further application.
Q7. What is the metric used to determine the quality of the retinal image?
Responses are aggregated into histograms, and a classifier is trained on these histograms’ counts in order to decide if a retinal image contains a reasonable visible proportion of such structures, under the assumption that the lack of presence of one of these clusters is an indicator of low quality, see [33] for the technical details.
Q8. How is the discriminator trained to classify codes z?
This is achieved via the maximization of the classification error of the discriminator module Dcode, which is trained to classify codes z sampled from q(z) according to whether they come from the true prior distribution p(z) or not.
Q9. How does the adversarial autoencoder achieve this?
The vessel-to-retinal image model presented in section II-A can map a vessel tree v to a realistic eye fundus image r, while the adversarial autoencoder defined in the previous section generates a vessel network v from a random sample z coming from a simple probability distribution.
Q10. How is the loss function used to maximize the classification error of the discriminator?
In addition, both the encoder and the decoder weights are updated to minimize the reconstruction error and, at the same time, to maximize the classification error of the discriminator.
Q11. How does the user generate a new pair of images?
the user is only required to sample from an N - dimensional predefined prior Gaussian distribution p(z) to generate a new pair of images.
Q12. How did the model achieve a 0.9755 AUC on the DRIVE test set?
This model achieved a 0.9755 AUC on the DRIVE test set, a result aligned with state-of-theart methods for retinal vessel segmentation [21]–[24].
Q13. How many pairs of images were used in this study?
For this reason, only images from Messidor-1 with grades 0, 1 and 2 were used in this work, reducing the number of example pairs to 946.
Q14. What is the metric used to report a quantitative image quality analysis?
to report a quantitative image quality analysis, the authors employ the Image Structure Clustering (ISC) metric proposed in [33].
Q15. What is the name of the dataset used for training?
1) The set of images used during training, containing 614 real retinal images and corresponding vessel trees extracted from the Messidor1 database, denoted Training Real Dataset (TrainRD); 2) The held-out test set, that was not used during training, and contains 177 real retinal images and associated vessel trees, denoted Test Real Dataset (TestRD).
Q16. How is the complete loss function that drives the learning of the adversarial autoencoder?
In this way, the complete loss function that drives the learning of the adversarial autoencoder is a combination of both losses:LAAE(Dcode, q, p) = Lcode(Dcode, q) + γLrec(q, p), (4)where γ weights the importance of the two losses.