1. What are the contributions mentioned in the paper "Multimodal mr synthesis via modality-invariant latent representation" ?
The authors propose a multi-input multi-output fully convolutional neural network model for MRI synthesis.. The authors also show that by incorporating information from segmentation masks the model can both decrease its error and generate data with synthetic lesions.. The authors evaluate their model on the ISLES and BRATS datasets and demonstrate statistically significant improvements over state-of-the-art methods for single input tasks.. Lastly, the authors demonstrate their approach on non skull-stripped brain images, producing a statistically significant improvement over the previous best method.. This improvement increases further when multiple input modalities are used, demonstrating the benefits of learning a common latent space, again resulting in a statistically significant improvement over the current best method.
read more
2. What are the evaluation metrics used to evaluate the performance of the methods?
To evaluate the performance of the methods, the authors use mean squared error (MSE), structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR).
read more
3. Why is it possible for very different latent representations to be decoded into very similar?
due to the highly non-linear, non-injective nature of the decoder, it is possible for very different latent representations (i.e. ones with a large Euclidean distance between them) to be decoded into very similar output images.
read more
4. What is the reason for the difficulty in synthesising non skull-stripped volumes?
As also discussed in [9], synthesising non skull-stripped volumes is difficult because of the intensity inhomogeneity in MR images caused by the dark skull regions surrounded by bright skin and fat regions.
read more
