Skip to content

Commit

Permalink
latest state of tumor maximisation
Browse files Browse the repository at this point in the history
  • Loading branch information
@janicweber
janicweber committed Apr 28, 2023
1 parent c778eab commit fc8affd
Showing 1 changed file with 35 additions and 34 deletions.
69 changes: 35 additions & 34 deletions src/adaptfx/reinforce.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -238,17 +238,17 @@ def min_n_frac(keys, sets=afx.SETTING_DICT):
output.policy = {}
output.policy.val = policy
output.policy.sf = sf
output.policy.states = bedt_states[::-1]
output.policy.states = bedt_states
if values_plot:
output.value = {}
output.value.val = values
output.value.sf = sf
output.value.states = bedt_states[::-1]
output.value.states = bedt_states
if remains_plot:
output.remains = {}
output.remains.val = remains
output.remains.sf = sf
output.remains.states = bedt_states[::-1]
output.remains.states = bedt_states

return output

Expand Down Expand Up @@ -322,18 +322,25 @@ def max_tumor_bed(keys, sets=afx.SETTING_DICT):
max_dose = min(max_dose, remaining_bed)

# actionspace in bed dose
bedn_space = np.linspace(0, remaining_bed, n_bedsteps + 1)
actionspace = afx.convert_to_physical(bedn_space, abn, actual_sf)
range_action = (bedn_space >= min_dose) & (bedn_space <= max_dose)
# bed_space to relate actionspace to oar- and tumor-dose
bedn_space = bedn_space[range_action]
actionspace = actionspace[range_action]
if not range_action.any():
# pre calculation
bedn_space_pre = np.linspace(0, remaining_bed, n_bedsteps + 1)
actionspace_pre = afx.convert_to_physical(bedn_space_pre, abn, actual_sf)
bedt_space_pre = afx.bed_calc0(actionspace_pre, abt)
range_action = (bedn_space_pre >= min_dose) & (bedn_space_pre <= max_dose)
if range_action.any():
# check if actionspace is empty
bedn_space = np.array([min_dose])
actionspace = afx.convert_to_physical(bedn_space, abn, actual_sf)
bedt_space = afx.bed_calc0(actionspace, abt)
n_action = len(actionspace)
bedn_space = bedn_space_pre[range_action]
actionspace = actionspace_pre[range_action]
bedt_space = bedt_space_pre[range_action]
n_action = len(bedn_space)
bedn_space_sf = np.zeros((1, n_sf)) + bedn_space.reshape(n_action, 1)
# bed_space to relate actionspace to oar- and tumor-dose
actionspace_sf = afx.convert_to_physical(bedn_space.reshape(n_action, 1), abn, sf)
else:
bedn_space_sf = np.zeros((1, n_sf)) + np.array([min_dose])
actionspace_sf = afx.convert_to_physical(bedn_space_sf, abn, sf)
n_action = 1
bedt_space_sf = afx.bed_calc0(actionspace_sf, abt)

# tumor bed states for tracking dose
oar_upper_bound = oar_limit + sets.state_stepsize
Expand All @@ -344,10 +351,6 @@ def max_tumor_bed(keys, sets=afx.SETTING_DICT):
remaining_states = bedn_states - accumulated_oar_dose
n_bedn_states = len(bedn_states)

# relate actionspace to bed and possible sparing factors
# necessary reshape for broadcasting in value calculation
actionspace_sf = afx.convert_to_physical(bedn_space.reshape(n_action, 1), abn, sf.reshape(1, n_sf))
bedt_sf_space = afx.bed_calc0(actionspace_sf, abt).reshape(1, n_action, n_sf)
# values matrix
# dim(values) = dim(policy) = fractions_remaining * bedt * sf
n_remaining_fractions = number_of_fractions - fraction
Expand Down Expand Up @@ -382,11 +385,11 @@ def max_tumor_bed(keys, sets=afx.SETTING_DICT):

if policy_plot or values_plot or remains_plot:
# for the policy plot
future_bedn_full = bedn_states.reshape(n_bedn_states, 1) + bedn_space
future_bedn_full = bedn_states.reshape(n_bedn_states, 1, 1) + bedn_space_sf.reshape(1, n_action, n_sf)
future_bedn_full = np.where(future_bedn_full > oar_limit, oar_upper_bound, future_bedn_full)
c_penalties_full = np.where(future_bedn_full < oar_limit, -c, 0).reshape(n_bedn_states, n_action, 1)
c_penalties_full = np.where(future_bedn_full < oar_limit, -c, 0)
future_values_full = afx.interpolate(future_bedn_full, bedn_states, future_values_discrete)
vs_full = bedt_sf_space + future_values_full.reshape(n_bedn_states, n_action, 1) + c_penalties_full
vs_full = bedt_space_sf.reshape(1, n_action, n_sf) + future_values_full + c_penalties_full
# check vs along the sf axis
current_policy = bedn_space[vs_full.argmax(axis=1)]
# ensure that for the goal reached the value/policy is zero (min_dose)
Expand Down Expand Up @@ -416,12 +419,11 @@ def max_tumor_bed(keys, sets=afx.SETTING_DICT):
# cut the actionspace to min and max dose constraints
last_bed_actions = np.where(last_bed_actions < min_dose, min_dose, last_bed_actions)
last_bed_actions = np.where(last_bed_actions > max_dose, max_dose, last_bed_actions)
last_actions = afx.convert_to_physical(last_bed_actions.reshape(n_bedn_states, 1), abn,
sf.reshape(1, n_sf))
last_bed_actions_sf = np.zeros((1, n_sf)) + last_bed_actions.reshape(n_bedn_states, 1)
last_bed_actions_reshaped = last_bed_actions.reshape(n_bedn_states, 1)
last_actions = afx.convert_to_physical(last_bed_actions_reshaped, abn, sf)
last_bedt = afx.bed_calc0(last_actions, abt)
# this smooths out the penalties in underdose and overdose regions
bedn_diff = np.round(bedn_states.reshape(n_bedn_states, 1) + last_bed_actions_sf - oar_limit, -exp)
bedn_diff = np.round(bedn_states.reshape(n_bedn_states, 1) + last_bed_actions_reshaped - oar_limit, -exp)
penalties = np.where(bedn_diff == 0, 0, -np.abs(bedn_diff) * sets.inf_penalty)
# to each best action add the according penalties
# penalties need to be reshaped for broadcasting
Expand All @@ -432,7 +434,7 @@ def max_tumor_bed(keys, sets=afx.SETTING_DICT):

if policy_plot:
# policy calculation for each bedt, but sf is not considered
policy[fraction_index] = last_bed_actions_sf
policy[fraction_index] += last_bed_actions_reshaped
# ensure that for the goal reached the value/policy is zero (min_dose)
policy[fraction_index][bedn_states==oar_limit] = 0

Expand All @@ -441,13 +443,13 @@ def max_tumor_bed(keys, sets=afx.SETTING_DICT):
# this calculates the value function in the future fractions
future_values_discrete = (values[fraction_index + 1] * prob).sum(axis=1)
# bedt_states is reshaped such that numpy broadcast leads to 2D array
future_bedn = bedn_states.reshape(n_bedn_states, 1) + bedn_space.reshape(1, n_action)
future_bedn = bedn_states.reshape(n_bedn_states, 1, 1) + bedn_space_sf.reshape(1, n_action, n_sf)
future_bedn = np.where(future_bedn > oar_limit, oar_upper_bound, future_bedn)
future_values = afx.interpolate(future_bedn, bedn_states, future_values_discrete)
c_penalties = np.where(future_bedn < oar_limit, -c, 0)
# dim(bedn_sf_space)=(1,n_action,n_sf),dim(future_values)=(n_states,n_action)
# every row of values_penalties is transposed and copied n_sf times
vs = bedt_sf_space + future_values.reshape(n_bedn_states, n_action, 1) + c_penalties.reshape(n_bedn_states, n_action, 1)
vs = bedt_space_sf.reshape(1, n_action, n_sf) + future_values + c_penalties
# check vs along the sf axis
values[fraction_index] = vs.max(axis=1)
# ensure that for the goal reached the value/policy is zero (min_dose)
Expand All @@ -466,10 +468,9 @@ def max_tumor_bed(keys, sets=afx.SETTING_DICT):
remains[fraction_index] = current_remains + future_remains

output = afx.DotDict({})

output.physical_dose = actionspace[action_index] if not finished else np.nan
output.tumor_dose = bedn_space[action_index] if not finished else np.nan
output.oar_dose = bedt_space[action_index] if not finished else np.nan
output.tumor_dose = bedt_space[action_index] if not finished else np.nan
output.oar_dose = bedn_space[action_index] if not finished else np.nan

if probability_plot:
output.probability = {}
Expand All @@ -479,16 +480,16 @@ def max_tumor_bed(keys, sets=afx.SETTING_DICT):
output.policy = {}
output.policy.val = policy
output.policy.sf = sf
output.policy.states = bedn_states[::-1]
output.policy.states = bedn_states
if values_plot:
output.value = {}
output.value.val = values
output.value.sf = sf
output.value.states = bedn_states[::-1]
output.value.states = bedn_states
if remains_plot:
output.remains = {}
output.remains.val = remains
output.remains.sf = sf
output.remains.states = bedn_states[::-1]
output.remains.states = bedn_states

return output

0 comments on commit fc8affd

Please sign in to comment.