diff --git a/src/aspire/classification/averager2d.py b/src/aspire/classification/averager2d.py
index 5db97f1aa3..9a91b77949 100644
--- a/src/aspire/classification/averager2d.py
+++ b/src/aspire/classification/averager2d.py
@@ -466,6 +466,8 @@ def __init__(
         if self.alignment_src.dtype != src.dtype:
             raise RuntimeError("Currently `alignment_src.dtype` must equal `src.dtype`")
 
+        self.mask = grid_2d(src.L, normalized=False)["r"] < src.L // 2
+
         super().__init__(composite_basis, src, composite_basis, dtype=dtype)
 
     def _phase_cross_correlation(self, img0, img1):
@@ -547,7 +549,7 @@ def _reddychatterji(self, images, class_k, reflection_k):
         # Result arrays
         M = len(images)
         rotations_k = np.zeros(M, dtype=self.dtype)
-        correlations_k = np.zeros(M, dtype=self.dtype)
+        correlations_k = np.full(M, -np.inf, dtype=self.dtype)
         shifts_k = np.zeros((M, 2), dtype=int)
 
         # De-Mean, note images is mutated and should be a `copy`.
@@ -655,8 +657,8 @@ def _reddychatterji(self, images, class_k, reflection_k):
             # Hack
             regis_img_estimated = rotate(regis_img, r)
             regis_img_rotated_p180 = rotate(regis_img, r + 180)
-            da = np.dot(fixed_img.flatten(), regis_img_estimated.flatten())
-            db = np.dot(fixed_img.flatten(), regis_img_rotated_p180.flatten())
+            da = np.dot(fixed_img[self.mask], regis_img_estimated[self.mask])
+            db = np.dot(fixed_img[self.mask], regis_img_rotated_p180[self.mask])
             if db > da:
                 regis_img_estimated = regis_img_rotated_p180
                 r += 180
@@ -702,7 +704,7 @@ def _reddychatterji(self, images, class_k, reflection_k):
                 shift = None  # For logger line
 
             # Estimated `corr` metric
-            corr = np.dot(fixed_img.flatten(), regis_img_estimated.flatten())
+            corr = np.dot(fixed_img[self.mask], regis_img_estimated[self.mask])
             correlations_k[m] = corr
 
             logger.debug(
@@ -962,9 +964,7 @@ def align(self, classes, reflections, basis_coefficients):
         L = self.alignment_src.L
 
         # Instantiate matrices for inner loop, and best results.
-        _rotations = np.zeros(classes.shape, dtype=self.dtype)
         rotations = np.zeros(classes.shape, dtype=self.dtype)
-        _correlations = np.zeros(classes.shape, dtype=self.dtype)
         correlations = np.ones(classes.shape, dtype=self.dtype) * -np.inf
         shifts = np.zeros((*classes.shape, 2), dtype=int)
 
@@ -986,16 +986,16 @@ def align(self, classes, reflections, basis_coefficients):
                 # Don't shift the base image
                 images[1:] = Image(unshifted_images[1:]).shift(s).asnumpy()
 
-                rotations[k], _, correlations[k] = self._reddychatterji(
+                _rotations, _, _correlations = self._reddychatterji(
                     images, classes[k], reflections[k]
                 )
 
                 # Where corr has improved
                 #  update our rolling best results with this loop.
-                improved = _correlations > correlations
-                correlations = np.where(improved, _correlations, correlations)
-                rotations = np.where(improved, _rotations, rotations)
-                shifts = np.where(improved[..., np.newaxis], s, shifts)
+                improved = _correlations > correlations[k]
+                correlations[k] = np.where(improved, _correlations, correlations[k])
+                rotations[k] = np.where(improved, _rotations, rotations[k])
+                shifts[k] = np.where(improved[..., np.newaxis], s, shifts[k])
                 logger.debug(f"Shift {s} has improved {np.sum(improved)} results")
 
         return rotations, shifts, correlations