Updated jem (#31)

Updated filters, features, and cli to include functions to generate feature set used for FCD paper

jem/cli.py

@@ -12,6 +12,7 @@
     fef_rotational_invariants,
     NUM_SCALES,
     NORMALIZATION_SCALE,
+    gaussian_pyramid_features
 )
 
 
@@ -158,7 +159,7 @@ def compute_laplacian_pyramid(input_image, num_scales, normalization_scale, outp
 )
 @click.option("--lowpass/--no-lowpass", default=False)
 def riff(input_image, num_scales, normalization_scale, output, lowpass):
-    """Compute rotationally invariant features."""
+    """Compute rotationally invariant bandpass features."""
 
     click.echo(f"Compute rotationally invariant features for {input_image}.")
 
@@ -210,4 +211,47 @@ def riff(input_image, num_scales, normalization_scale, output, lowpass):
     out_im.set_data_dtype(np.float32)
     out_im.to_filename(output)
 
-    click.echo(f"Wrote rotationally invariant features to {output}.")
+    click.echo(f"Wrote rotationally invariant bandpass features to {output}.")
+
+@click.command()
+@click.option(
+    "--output",
+    type=click.STRING,
+    default="out.nii",
+    help="Output filename for the features image.",
+)
+@click.argument("input_image", type=click.STRING)
+@click.option(
+    "--num_scales", type=click.INT, default=NUM_SCALES, help="number of spatial scales"
+)
+@click.option(
+    "--normalization_scale",
+    type=click.INT,
+    default=NORMALIZATION_SCALE,
+    help="scale for input gain control",
+)
+def compute_features(input_image, num_scales, normalization_scale, output):
+    """Compute rotationally invariant features."""
+
+    click.echo(f"Compute rotationally invariant features for {input_image}.")
+
+    # open the images
+    im = nibabel.load(input_image)
+    data = im.get_data().astype(np.float32)
+
+    # Global scaling, signal likelihood, noise level
+    f, w, sigma = global_scale(data)
+
+    # Local scale normalization
+    f = local_scale_normalization(
+        f, w=w, sigma=sigma, normalization_scale=normalization_scale
+    )
+
+    # compute features
+    feats = gaussian_pyramid_features(f, num_scales=num_scales, w=w, sigma=sigma)
+    feats = np.stack(feats, axis=-1)
+
+    # write out the result in the same format and preserve the header
+    out_im = type(im)(feats, affine=None, header=im.header)
+    out_im.set_data_dtype(np.float32)
+    out_im.to_filename(output)

jem/features.py

@@ -1,6 +1,6 @@
 import numpy as np
 
-from .filters import _pinv, dog, dog_rotational_invariants
+from .filters import _pinv, dog, dog_rotational_invariants, gradient_amplitude, hessian_amplitude, hessian_trace
 
 # Number of spatial scales
 NUM_SCALES = 4
@@ -215,3 +215,17 @@ def fef_rotational_invariants(fef, inplace=True):
         rfef["two"][n] = dog_rotational_invariants(fef["two"][n], order=2)
 
     return rfef
+
+def gaussian_pyramid_features(d, w, sigma, num_scales=NUM_SCALES):
+    """Rotational invariants of the gaussian pyramid features
+    """
+
+    gauss = gaussian_pyramid(d, order=0, num_scales=num_scales, w=w, sigma=sigma)
+    grad = gaussian_pyramid(d, order=1, num_scales=num_scales, w=w, sigma=sigma)
+    grad = [gradient_amplitude(x) for x in grad]
+    hess = gaussian_pyramid(d, order=2, num_scales=num_scales, w=w, sigma=sigma)
+    lap = [hessian_trace(x) for x in hess]
+    norm = [hessian_amplitude(x) for x in hess]
+    features = [d]+gauss+grad+lap+norm
+
+    return features

jem/filters.py

@@ -438,6 +438,21 @@ def hessian_amplitude(h):
         raise RuntimeError("Unsupported number of dimensions {}.".format(len(h)))
     return a
 
+def hessian_det(h):
+    """
+    Determinant in the hessian filter band
+    """
+    if len(h) == 3:
+        det = h[0] * h[2] - h[1] * h[1]
+    elif len(h) == 6:
+        det = (
+            h[0] * (h[3] * h[5] - h[4] * h[4])
+            - h[1] * (h[1] * h[5] - h[4] * h[2])
+            + h[2] * (h[1] * h[4] - h[3] * h[2])
+        )
+    else:
+        raise RuntimeError("Unsupported number of dimensions {}.".format(len(h)))
+    return det
 
 def hessian_rot(h):
     """
@@ -494,6 +509,18 @@ def hessian_rot(h):
     else:
         raise RuntimeError("Unsupported number of dimensions {}.".format(len(h)))
 
+def hessian_trace(h):
+    """
+    Trace in the hessian filter band
+    Laplacian
+    """
+    if len(h) == 3:
+        trace = h[0] + h[2]
+    elif len(h) == 6:
+        trace = h[0] + h[3] + h[5]
+    else:
+        raise RuntimeError("Unsupported number of dimensions {}.".format(len(h)))
+    return trace
 
 def rotate_gradient_2d(gx, gy, z=0.0):
     """Rotate a 2d gradient or band pass gradient.

setup.py

@@ -55,6 +55,7 @@
             "coil_correction=jem.cli:coil_correction",
             "laplacian_pyramid=jem.cli:compute_laplacian_pyramid",
             "riff=jem.cli:riff",
+            "compute_features=jem.cli:compute_features",
         ]
     },
     include_package_data=True,