Week #22
25 Mar 2019Experiments Results
| Gourp Id | Parameter num | initial lr | step_size | gamma | epoch | Loss |
| 0 | 13177026 | 0.001 | 100 | 0.8 | 350 | 0.0402 |
| 1 | 13177026 | 0.001 | 100 | 0.85 | 750 | 0.1314 |
| 2 | 13177026 | 0.001 | 100 | 0.85 | 2000 | 0.3422 |
| 3 | 13177026 | 0.001 | 100 | 0.85 | 2000 | 0.9754 |
| 3 | 52633794 | 0.001 | 100 | 0.85 | 2000 | 0.2850521 |
One possible solution
Inspired from R-CNN which is used for object detection and semantic segmentation at the same time. We could consider the first scale parcellation as the bounding box for our MRI images. And in each bounding box, the post-process could be down using the FCNN (Fully convolutional neural network). By this way, the first level parcellation could be used for retrieving the location information, which could decrease the computational expenses compared to one single image segmentation process.
The structure is showed as below:
Paper Construction
