appliedai-initiative / armscan_env Goto Github PK

So that the reward is not calculated multiple times for the observations, it is best to set a @cache call.
The idea is to use @lru_cache(max_size=100).
This means, that the parameters of the function anatomy_based_rwd shall be hashable.
Its parameters are tissue_clusters and n_landmarks.

PArameter tissue_clusters is a dataclass TissuClusters which consists as well of the dataclass elements DataCluster. In this two dataclasses, the __hash__ method needs to be implemented.

Test for each volume that for any random volume transformation, the optimal action is transformed accordingly.

Change the extent of the images, so that they are shown in metric dimension instead of pixel indexes.

Pad the volumes with 0s so that the arrays have same dimension when rendering., keeping the model in the hand.

All volumes should have the same spacing, hence resolution.
The size of the volumes should be changed accordingly, so that:
new_size = physical_size / spacing

Take the cluster size into account when calculating the reward

Test that for each volume, its optimal slice action leads to a reward smaller than termination threshold.

Volume optimal action is a fixed property which should not be able to change.
Running the env_rollout function in L5 notebook seems to modify it.

Since the volume orientation is saved under GetDirection(), the random rotation applied do augment the data does not have any effect on the slicing actions. In fact, it is possible to get the same slice without changing the action rotation, even though the volume has been transformed with a random rotation.