Merge pull request #27 from jagruti8/add-detectors-metrics

Added new outlier detector methods, modified the README.md file and modified the algorithm.txt

README.md

@@ -114,24 +114,28 @@ identified as an outlier.  0 refers to the first volume.  For example (these
 outlier IDs are completely random, for illustration):
 
 ```
-data/group-01/sub-08/func/sub-08_task-taskzero_run-01_bold.nii.gz, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 18, 19, 133, 134, 135, 136, 154, 155, 157
-data/group-01/sub-08/func/sub-08_task-taskzero_run-02_bold.nii.gz, 0, 1, 2, 3, 4, 5, 6, 9, 17, 53, 54, 63, 78, 79, 151, 152, 153
-data/group-01/sub-01/func/sub-01_task-taskzero_run-01_bold.nii.gz, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 20, 21, 157, 158
-data/group-01/sub-01/func/sub-01_task-taskzero_run-02_bold.nii.gz, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 14, 17, 19, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160
-data/group-01/sub-06/func/sub-06_task-taskzero_run-02_bold.nii.gz, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 153, 154, 155, 156, 157, 158, 159, 160, 161
-data/group-01/sub-06/func/sub-06_task-taskzero_run-01_bold.nii.gz, 0, 1, 2, 3, 4, 5, 6, 7, 8, 19, 24, 25, 26, 27, 28, 29, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159
-data/group-01/sub-07/func/sub-07_task-taskzero_run-02_bold.nii.gz, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28
-data/group-01/sub-07/func/sub-07_task-taskzero_run-01_bold.nii.gz, 0, 1, 2, 3, 4, 5, 6, 7, 132, 136, 137, 138, 139, 140, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161
-data/group-01/sub-09/func/sub-09_task-taskzero_run-01_bold.nii.gz, 0, 134, 135, 136, 143, 144, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160
-data/group-01/sub-09/func/sub-09_task-taskzero_run-02_bold.nii.gz, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 36, 79, 80, 150, 151, 152, 153, 154, 155, 156, 157
-data/group-01/sub-10/func/sub-10_task-taskzero_run-02_bold.nii.gz, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 26, 104
-data/group-01/sub-10/func/sub-10_task-taskzero_run-01_bold.nii.gz, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159
-data/group-01/sub-05/func/sub-05_task-taskzero_run-01_bold.nii.gz, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 20, 25, 26, 27, 48, 49, 52, 76, 77, 150
-data/group-01/sub-05/func/sub-05_task-taskzero_run-02_bold.nii.gz, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 28, 50, 51, 52, 54, 157, 158, 159, 160, 161
-data/group-01/sub-02/func/sub-02_task-taskzero_run-02_bold.nii.gz, 34, 65, 105, 106, 107, 135, 140, 148
-data/group-01/sub-02/func/sub-02_task-taskzero_run-01_bold.nii.gz, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21
-data/group-01/sub-03/func/sub-03_task-taskzero_run-02_bold.nii.gz, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 101, 102, 103, 160, 161
-data/group-01/sub-03/func/sub-03_task-taskzero_run-01_bold.nii.gz, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 14, 15, 136, 137, 138, 139, 140, 142, 156, 157, 158, 159
-data/group-01/sub-04/func/sub-04_task-taskzero_run-01_bold.nii.gz, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 12, 59, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161
-data/group-01/sub-04/func/sub-04_task-taskzero_run-02_bold.nii.gz, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 32, 33, 34, 35, 36, 49, 50, 52, 53, 54, 55, 57, 58, 148, 149, 157
-```
+data/group-01/sub-08/func/sub-08_task-taskzero_run-01_bold.nii.gz, 11, 157
+data/group-01/sub-08/func/sub-08_task-taskzero_run-02_bold.nii.gz, 79, 153
+data/group-01/sub-01/func/sub-01_task-taskzero_run-01_bold.nii.gz, 0, 153
+data/group-01/sub-01/func/sub-01_task-taskzero_run-02_bold.nii.gz, 151
+data/group-01/sub-06/func/sub-06_task-taskzero_run-02_bold.nii.gz, 0, 1, 21, 22, 23, 24, 25, 26, 28, 29, 155
+data/group-01/sub-06/func/sub-06_task-taskzero_run-01_bold.nii.gz, 1
+data/group-01/sub-07/func/sub-07_task-taskzero_run-02_bold.nii.gz, 79, 80
+data/group-01/sub-07/func/sub-07_task-taskzero_run-01_bold.nii.gz, 85
+data/group-01/sub-09/func/sub-09_task-taskzero_run-02_bold.nii.gz, 23, 24, 25, 26, 27, 28, 30
+data/group-01/sub-10/func/sub-10_task-taskzero_run-02_bold.nii.gz, 104
+data/group-01/sub-05/func/sub-05_task-taskzero_run-01_bold.nii.gz, 0, 49, 77, 150
+data/group-01/sub-05/func/sub-05_task-taskzero_run-02_bold.nii.gz, 3, 4, 5, 6, 9, 20, 23, 28, 49, 54
+data/group-01/sub-03/func/sub-03_task-taskzero_run-02_bold.nii.gz, 160, 161
+data/group-01/sub-03/func/sub-03_task-taskzero_run-01_bold.nii.gz, 11, 14, 15, 132, 156, 157
+data/group-01/sub-04/func/sub-04_task-taskzero_run-01_bold.nii.gz, 1, 144, 154, 158, 159, 160, 161
+data/group-01/sub-04/func/sub-04_task-taskzero_run-02_bold.nii.gz, 0, 1, 28, 35, 49, 52, 53
+```
+
+Shown below are the plots of the mean of voxel intensities for each time point vs time points. The detected outliers are marked in orange colour:
+![](sub_01_02.png)
+![](sub_03_04.png)
+![](sub_05_06.png)
+![](sub_07_08.png)
+![](sub_09_10.png)
+

algorithm.txt

@@ -1,10 +1,28 @@
 Algorithm:
 
 1. The 4D image is first segmented (using otsu threshold) to segment the brain voxels from the background. (x)
+
+I. Median absolute deviation over voxels and median absolute deviation over time points
 2. The median(med_voxel(x)) and median absolute deviation(mad_voxel(x)) is calculated for each of the brain voxels.
 3. The brain voxels lying outside the interval [med_voxel(x)-a*mad_voxel(x), med_voxel(x)+a*mad_voxel(x)] are considered as outliers. a = 3.5
 4. For each time t, the number of outlying voxels n(t) is counted.
 5. The median (n_med) and MAD (n_mad) of n(t) are calculated. Any time t with n(t)>n_med+3.5*n_mad are considered as outliers.
 
+II. DVARS and median absolute deviation over time points
+6. The dvars(t) of the brain voxels are calculated.
+7. The median (dvars_med) and MAD (dvars_mad) of dvars(t) are calculated. Any time t with |dvars(t)-dvars_med|>3.5*dvars_mad are considered as outliers.
+
+III. Sliding window and median absolute deviation over time points
+8. A fraction of the time points are chosen using a sliding window, mean over voxel intensities for each time point (m(t)) in this sliding window is calculated. 
+9. The median (m_med) and MAD (m_mad) of m(t) in this sliding window are calculated. Any time t with |m(t)-m_med|>3.5*m_mad are considered as outliers. 
+10. This is repeated till the sliding window covers all the time points.
+11. All the outliers in each sliding window are merged.
+
+12. The outliers from the I approach is a more global approach and filters a lot of time-points as outliers. 
+13. The outliers from the II(DVARS) approach compares successive volumes and if the difference is large considers the preceding volumes as outliers.
+14. The outliers from the III(sliding window) approach compares volumes within a certain range and then detect outliers. This is mostly done to take care of drift.
+
+15. Outliers from I and II are merged (o_total) (Global + Local). If these o_total agree with the outliers from sliding window, they are classified as final outliers (o_final).
+
 References:
 1. Cox, R.W. Outlier Detection in FMRl Time Series. ISMRM(2002).

findoutlie/detectors.py

@@ -45,7 +45,7 @@ def mad_voxel_detector(img,threshold=3.5):
     outlier_tf = np.abs(img-med)>(threshold*mad)
     return outlier_tf
 
-def mad_time_detector(measures, threshold=3.5):
+def mad_time_detector(measures, lower_bound, threshold=3.5):
     """ Detect outliers in 'measures' using median absolute deviation.
     Returns 1D vector of same length as 'measures', where True means the corresponsding 
     value in 'measures' is an outlier.
@@ -71,7 +71,10 @@ def mad_time_detector(measures, threshold=3.5):
     # Calculate median absoulte deviation of measures
     mad = np.median(np.abs(measures-med))
     # Calculate the outliers
-    outlier_tf = measures>med+threshold*mad
+    if lower_bound:
+        outlier_tf = np.abs(measures-med)>threshold*mad
+    else:
+        outlier_tf = measures>med+threshold*mad
     return outlier_tf
 
 def iqr_detector(measures, iqr_proportion=1.5):

findoutlie/metrics.py

@@ -5,6 +5,26 @@
 # +++your code here+++
 import numpy as np
 
+def dvars_voxel(voxels):
+    """ Calculate dvars metric on 2D array with voxels in rows and time-points(volumes) in columns
+
+    The dvars calculation between two volumes is defined as the square root of 
+    (the mean of the (voxel differences square)).
+
+    Parameters
+    ----------
+    voxels : 2D array
+
+    Returns
+    -------
+    dvals : 1D array
+        One-dimensional array with n-1 elements, where n is the number of 
+        volumes in 'img'.
+    """
+    vol_diff = voxels[..., 1:] - voxels[..., :-1]
+    dvar_val = np.sqrt(np.mean(vol_diff ** 2, axis=0))
+    return dvar_val
+
 
 def dvars(img):
     """ Calculate dvars metric on Nibabel image `img`
@@ -33,14 +53,8 @@ def dvars(img):
     data = img.get_fdata()
 
     voxel_by_time = np.reshape(data, (-1, data.shape[-1]))
+    dvar_val = dvars_voxel(voxel_by_time)
 
-    vol_diff = voxel_by_time[..., 1:] - voxel_by_time[..., :-1]  # 2D array
-    # print(vol_diff.shape())
-    # print(vol_diff)
-    # vol_diff_1D=vol_diff.flatten()
-    dvar_val = np.sqrt(np.mean(vol_diff ** 2, axis=0))
-    # print(dvar_val.shape())
-    # print(dvar_val)
     return dvar_val
 
     raise NotImplementedError("Code up this function")

findoutlie/outfind.py

@@ -9,7 +9,7 @@
 
 from skimage.filters import threshold_otsu
 
-from .metrics import dvars
+from .metrics import dvars,dvars_voxel
 from .detectors import iqr_detector,mad_voxel_detector,mad_time_detector
 
 def segment_brain(img):
@@ -27,14 +27,12 @@ def segment_brain(img):
     mean_img = np.mean(img, axis=-1)
     # calculate the threshold for segmenting brain from background
     threshold = threshold_otsu(mean_img)
-    mask = np.expand_dims(mean_img > threshold, axis=1)
-    mask_2D = np.tile(mask, (1, img.shape[-1]))
-    thresholded_img = np.where(mask_2D, img, np.nan)
+    mask = mean_img > threshold
     # filter only brain voxels
-    brain_voxels = thresholded_img[~np.isnan(thresholded_img).all(axis=1)]
+    brain_voxels = img[mask]
     return brain_voxels
 
-def detect_outliers_mean_absolute_deviation_mask(fname):
+def detect_outliers_mad_median_absolute_deviation_mask(fname):
     """ Detect outliers given image file path 'filename'
      
     Parameters
@@ -47,7 +45,7 @@ def detect_outliers_mean_absolute_deviation_mask(fname):
     outliers : array
         Indices of outlier volumes.
     """
-    # A mask is used to first segment the brain regions from the background, then mean absolute deviation is used to detect outliers
+    # A mask is used to first segment the brain regions from the background, then median absolute deviation is used to detect outliers
     img = nib.load(fname)
     img_data = img.get_fdata()
     # reshape from 4D to 2D
@@ -59,9 +57,71 @@ def detect_outliers_mean_absolute_deviation_mask(fname):
     # calculate the number of outlying voxels for each time point
     voxel_outliers_per_time = np.nansum(outliers_voxel,axis=0)
     # find the outliers in the time-series    
-    outliers_time = mad_time_detector(voxel_outliers_per_time)
+    outliers_time = mad_time_detector(voxel_outliers_per_time, lower_bound=False)
     # Return indices of True values from Boolean array. 
-    return np.nonzero(outliers_time)[0]
+    return np.nonzero(outliers_time)
+
+def detect_outliers_mad_dvars_mask(fname):
+    """ Detect outliers given image file path 'filename'
+    
+    Parameters
+    ----------
+    fname : str or Path
+        Filename of 4D image, as string or Path object
+    
+    Returns
+    -------
+    outliers : array
+        Indices of outlier volumes.
+    """
+    # A mask is used to first segment the brain regions from the background, dvars is calculated and then median absolute deviation is used to detect outliers 
+    img = nib.load(fname)
+    img_data = img.get_fdata()
+    # reshape from 4D to 2D
+    img_data_2D = np.reshape(img_data, (-1,img_data.shape[-1]))
+    # segment brain from background
+    brain_voxels = segment_brain(img_data_2D)
+    # calculate dvars
+    dvs = dvars_voxel(brain_voxels)
+    # detect outliers
+    is_outlier = mad_time_detector(dvs, lower_bound=True)
+    return np.nonzero(is_outlier)
+
+def detect_outliers_mad_sliding_window_mask(fname):
+    """ Detect outliers given image file path 'filename'
+    
+    Parameters
+    ----------
+    fname : str or Path
+        Filename of 4D image, as string or Path object
+    
+    Returns 
+    -------
+    outliers : array
+        Indices of outlier volumes.
+    """
+    # A mask is used to first segment the brain regions from the background, sliding window approach is used to detect outliers in each window using median absolute deviation
+    img = nib.load(fname)
+    img_data = img.get_fdata()
+    # reshape from 4D to 2D
+    img_data_2D = np.reshape(img_data, (-1,img_data.shape[-1]))
+    # segment brain from background
+    brain_voxels = segment_brain(img_data_2D)
+    # apply sliding window
+    overlap = 10
+    window_length = 20
+    outliers = []
+    for i in range(0, brain_voxels.shape[-1],overlap):
+       if i+window_length>=brain_voxels.shape[-1]:
+           elements = np.mean(brain_voxels[:,i:],axis=0)
+           outliers_1 = np.nonzero(mad_time_detector(elements, lower_bound=True))[0] + i
+           outliers.extend(outliers_1)
+           break
+       else:
+           elements = np.mean(brain_voxels[:,i:i+window_length],axis=0)
+           outliers_1 = np.nonzero(mad_time_detector(elements, lower_bound=True))[0] + i
+           outliers.extend(outliers_1)
+    return np.unique(outliers)
 
 
 def detect_outliers(fname):
@@ -82,7 +142,7 @@ def detect_outliers(fname):
     dvs = dvars(img)
     is_outlier = iqr_detector(dvs, iqr_proportion=2)
     # Return indices of True values from Boolean array.
-    return np.nonzero(is_outlier)[0]
+    return np.nonzero(is_outlier)
 
 
 def find_outliers(data_directory):
@@ -102,7 +162,13 @@ def find_outliers(data_directory):
     image_fnames = Path(data_directory).glob("**/sub-*.nii.gz")
     outlier_dict = {}
     for fname in image_fnames:
-        outliers = detect_outliers_mean_absolute_deviation_mask(fname)
+        # detect outliers using mad over voxels and mad over time points
+        outliers_mad = detect_outliers_mad_median_absolute_deviation_mask(fname)
+        # detect outliers using dvars and mad over time points
+        outliers_dvars = detect_outliers_mad_dvars_mask(fname)
+        # detect outliers using sliding window and mad over time points
+        outliers_sliding_window = np.array(detect_outliers_mad_sliding_window_mask(fname))
+        outliers = np.intersect1d(outliers_sliding_window,np.union1d(outliers_mad, outliers_dvars))
         #outliers = detect_outliers(fname)
         outlier_dict[fname] = outliers
     return outlier_dict