Week 5 - Alexandra Roberts

 

Last week I evaluated the performance of a pre-trained Very Deep Super Resolution (VDSR) network on generating a 7T quality scan from a (downsampled 7T) 3T analog. This week I trained the network on 504 patches from 9 patient datasets collected from a 7T scanner, and included a de-noising operation on the residual. The network performance was tested on an unseen dataset and measured by calculating the mean squared error between the ground truth 7T and the VDSR output. This performance was compared to the traditional bicubic interpolation method, which calculates the upsampled image pixel value from a 4x4 neighborhood around each pixel of the low-resolution image. The VDSR output shows increased sharpness as compared to the bicubic interpolation.



In addition to sharpening the image, the VDSR (orange) outperforms the bicubic interpolation (blue) as measured by its lower mean squared error with the ground truth.



To improve the output quality as compared to bicubic interpolation, the network will be re-trained on actual 3T and 7T datasets. This should better capture the physical differences between 3T and 7T in the residual.


Comments

Popular posts from this blog

Week 1

Week 1 - Samantha Bratcher

Week One - Shenghao Tan