Doserad challange

Todos:

• Figure out what is the best spacing configuration
• Find out hierarchy of training ( beam level, ray level, beamlet level)

Training structure

Metadata encoding

Encoding extra metadata proton/SynthRad...

• One way to encode metadata or multiple models.
• What metadata to encode?
• Does it even help?

Encoding base metadata

• Should we use MLP encoding?
• Geometric field encoding. (Source and target encoding)
• Fourirer encoding.
• Postional encoding (sine and cosine)

Preprocess

Training level

• Should we train at beam level, ray level or beamlet level?

Data augmentation

• Should we train on HU or RSP?
• What is the ideal size of input volume? (spacing, target size)

Losses.

• Use only competition default loss?
• Use combinations?
• If latent transition, should we use extra regularizations?

Model architectures.

• Vision Transformer.
• Conv encoder + Transformer transition + Conv decoder.
• 3D Unet.

Planned Meetings:

• Wed: 13:00 - 14:00
• Fri: 13:00 - 14:00