Project Week 25/CNN for PseudoCT Generation from T1T2 MR
Back to Projects List
- Giampaolo Pileggi (Magna Graecia University, Italy/German Cancer Research Center (DKFZ), Germany)
- Paolo Zaffino (Magna Graecia University, Italy)
- Salvatore Scaramuzzino (Magna Graecia University/ASL Vercelli, Italy)
- Maria Francesca Spadea (Magna Graecia University, Italy)
- Gino Gulamhussene (University of Magdeburg, Germany)
- Anneke Meyer (University of Magdeburg, Germany)
This tool allows the user to generate an HU map (Pseudo-CT) from T1/T2 input MRI. The tool is still in the early development stages with good early results in terms of Mean Absolute Error and Bias.
|Objective||Approach and Plan||Progress and Next Steps|
The main object is to understand which topology of CNN is the most suited for the Pseudo-CT generation task.
Test of different structures with training on low-res images in order to speed-up computational time. Analysis of the different output with MAE and Bias metrics
The most suited CNN to be used for this task consisted of an edited U-Net that does not make use of the final Dense Layer, useful for segmentation but not for regression purposes. Also unpooling layer was used in the decoding part, with zero-filling instead of repetitions for the missing values. Next planned steps will be the fine tuning of the CNN in order to lower the MAE and BIAS, the translation to TensorFlow (now the code uses Theano as a backend) and the data parallelization. Interactions with Annette, Gino and Gabriele were important in order to understand how to improve the CNN topology.