Further Additions Updates to BayesianNN

50-class-mapping.csv

@@ -0,0 +1,107 @@
+,original,new,label
+0,0,0,Unknown
+1,2,1,Left-Cerebral-White-Matter
+2,4,2,Left-Lateral-Ventricle
+3,5,2,Left-Inf-Lat-Vent
+4,7,3,Left-Cerebellum-White-Matter
+5,8,4,Left-Cerebellum-Cortex
+6,10,5,Left-Thalamus-Proper
+7,11,6,Left-Caudate
+8,12,7,Left-Putamen
+9,13,8,Left-Pallidum
+10,14,2,3rd-Ventricle
+11,15,2,4th-Ventricle
+12,16,9,Brain-Stem
+13,17,10,Left-Hippocampus
+14,18,11,Left-Amygdala
+15,24,12,CSF
+16,26,13,Left-Accumbens-area
+17,28,14,Left-VentralDC
+18,41,1,Right-Cerebral-White-Matter
+19,43,2,Right-Lateral-Ventricle
+20,44,2,Right-Inf-Lat-Vent
+21,46,3,Right-Cerebellum-White-Matter
+22,47,4,Right-Cerebellum-Cortex
+23,49,5,Right-Thalamus-Proper
+24,50,6,Right-Caudate
+25,51,7,Right-Putamen
+26,52,8,Right-Pallidum
+27,53,10,Right-Hippocampus
+28,54,11,Right-Amygdala
+29,58,13,Right-Accumbens-area
+30,60,14,Right-VentralDC
+31,72,2,5th-Ventricle
+32,192,15,Corpus_Callosum
+33,251,15,CC_Posterior
+34,252,15,CC_Mid_Posterior
+35,253,15,CC_Central
+36,254,15,CC_Mid_Anterior
+37,255,15,CC_Anterior
+38,1001,16,ctx-lh-bankssts
+39,1002,17,ctx-lh-caudalanteriorcingulate
+40,1003,18,ctx-lh-caudalmiddlefrontal
+41,1005,19,ctx-lh-cuneus
+42,1006,20,ctx-lh-entorhinal
+43,1007,21,ctx-lh-fusiform
+44,1008,22,ctx-lh-inferiorparietal
+45,1009,23,ctx-lh-inferiortemporal
+46,1010,24,ctx-lh-isthmuscingulate
+47,1011,25,ctx-lh-lateraloccipital
+48,1012,26,ctx-lh-lateralorbitofrontal
+49,1013,27,ctx-lh-lingual
+50,1014,28,ctx-lh-medialorbitofrontal
+51,1015,29,ctx-lh-middletemporal
+52,1016,30,ctx-lh-parahippocampal
+53,1017,31,ctx-lh-paracentral
+54,1018,32,ctx-lh-parsopercularis
+55,1019,33,ctx-lh-parsorbitalis
+56,1020,34,ctx-lh-parstriangularis
+57,1021,35,ctx-lh-pericalcarine
+58,1022,36,ctx-lh-postcentral
+59,1023,37,ctx-lh-posteriorcingulate
+60,1024,38,ctx-lh-precentral
+61,1025,39,ctx-lh-precuneus
+62,1026,40,ctx-lh-rostralanteriorcingulate
+63,1027,41,ctx-lh-rostralmiddlefrontal
+64,1028,42,ctx-lh-superiorfrontal
+65,1029,43,ctx-lh-superiorparietal
+66,1030,44,ctx-lh-superiortemporal
+67,1031,45,ctx-lh-supramarginal
+68,1032,46,ctx-lh-frontalpole
+69,1033,47,ctx-lh-temporalpole
+70,1034,48,ctx-lh-transversetemporal
+71,1035,49,ctx-lh-insula
+72,2001,16,ctx-rh-bankssts
+73,2002,17,ctx-rh-caudalanteriorcingulate
+74,2003,18,ctx-rh-caudalmiddlefrontal
+75,2005,19,ctx-rh-cuneus
+76,2006,20,ctx-rh-entorhinal
+77,2007,21,ctx-rh-fusiform
+78,2008,22,ctx-rh-inferiorparietal
+79,2009,23,ctx-rh-inferiortemporal
+80,2010,24,ctx-rh-isthmuscingulate
+81,2011,25,ctx-rh-lateraloccipital
+82,2012,26,ctx-rh-lateralorbitofrontal
+83,2013,27,ctx-rh-lingual
+84,2014,28,ctx-rh-medialorbitofrontal
+85,2015,29,ctx-rh-middletemporal
+86,2016,30,ctx-rh-parahippocampal
+87,2017,31,ctx-rh-paracentral
+88,2018,32,ctx-rh-parsopercularis
+89,2019,33,ctx-rh-parsorbitalis
+90,2020,34,ctx-rh-parstriangularis
+91,2021,35,ctx-rh-pericalcarine
+92,2022,36,ctx-rh-postcentral
+93,2023,37,ctx-rh-posteriorcingulate
+94,2024,38,ctx-rh-precentral
+95,2025,39,ctx-rh-precuneus
+96,2026,40,ctx-rh-rostralanteriorcingulate
+97,2027,41,ctx-rh-rostralmiddlefrontal
+98,2028,42,ctx-rh-superiorfrontal
+99,2029,43,ctx-rh-superiorparietal
+100,2030,44,ctx-rh-superiortemporal
+101,2031,45,ctx-rh-supramarginal
+102,2032,46,ctx-rh-frontalpole
+103,2033,47,ctx-rh-temporalpole
+104,2034,48,ctx-rh-transversetemporal
+105,2035,49,ctx-rh-insul

bayesian_utils.py

@@ -0,0 +1,20 @@
+import os
+import numpy as np
+import tensorflow as tf
+import tensorflow.keras.backend as K
+from tensorflow.keras.callbacks import Callback
+import tensorflow_probability as tfp
+import math
+
+def normal_prior(mu, prior_std):
+    """Defines distributions prior for Bayesian neural network.
+       Simply set tf.zeros(shape, dtype) with a new mu value for any new distribution  as needed
+       I tried Normal, He, Xavier. 
+    """
+    def prior_fn(dtype, shape, name, trainable, add_variable_fn):
+        tfd = tfp.distributions
+        dist = tfd.Normal(loc=tf.zeros(shape, dtype),
+                          scale=dtype.as_numpy_dtype((prior_std)))
+        batch_ndims = tf.size(input=dist.batch_shape_tensor())
+        return tfd.Independent(dist, reinterpreted_batch_ndims=batch_ndims)
+    return prior_fn

data_loader.py

@@ -0,0 +1,66 @@
+"""
+Created on Tue Feb 23 18:34:40 2021
+
+@author: aakanksha
+"""
+# Imports
+import nobrainer
+import tensorflow as tf
+import sys
+import json
+import glob
+import numpy as np
+import pandas as pd
+import os
+import warnings
+import nibabel as nib
+
+
+
+def _to_blocks(x, y,block_shape):
+    """Separate `x` into blocks and repeat `y` by number of blocks."""
+    print(x.shape)
+    x = nobrainer.volume.to_blocks(x, block_shape)
+    y = nobrainer.volume.to_blocks(y, block_shape)
+    return (x, y)
+
+def get_dict(n_classes):
+    print('Conversion into {} segmentation classes from freesurfer labels to 0-{}'.format(n_classes,n_classes-1))
+    if n_classes == 50: 
+        tmp = pd.read_csv('50-class-mapping.csv')
+        mydict = dict(tuple(zip(tmp['original'],tmp['new'])))
+        return mydict
+    elif n_classes == 115:
+        tmp = pd.read_csv('115-class-mapping.csv', header=0,usecols=[0,1],dtype=np.int32)
+        mydict = dict(tuple(zip(tmp['original'],tmp['new'])))
+        return mydict
+    else: raise(NotImplementedError)
+
+def process_dataset(dset,batch_size,block_shape,n_classes,train= True):
+    # Standard score the features.
+    dset = dset.map(lambda x, y: (nobrainer.volume.standardize(x), nobrainer.volume.replace(y,get_dict(n_classes))))
+    # Separate features into blocks.
+    dset = dset.map(lambda x, y:_to_blocks(x,y,block_shape))
+    # This step is necessary because separating into blocks adds a dimension.
+    dset = dset.unbatch()
+    #dset = dset.shuffle(buffer_size=100)
+    # Only shuffle the dset for training
+    if train:
+        dset = dset.shuffle(buffer_size=100)
+    else:
+        pass
+    # Add a grayscale channel to the features.
+    dset = dset.map(lambda x, y: (tf.expand_dims(x, -1), y))
+    # Batch features and labels.
+    dset = dset.batch(batch_size, drop_remainder=True)
+    return dset
+
+def get_dataset(pattern,volume_shape,batch,block_shape,n_classes,train=True):
+    dataset = nobrainer.dataset.tfrecord_dataset(
+        file_pattern=glob.glob(pattern),
+        volume_shape=volume_shape,
+        shuffle=False,
+        scalar_label=False,
+        compressed=True)
+    dataset = process_dataset(dataset,batch,block_shape,n_classes,train=train)
+    return dataset