diff --git a/lidar_pipeline/visualizations.py b/lidar_pipeline/visualizations.py
index 9ebb913..13e5390 100644
--- a/lidar_pipeline/visualizations.py
+++ b/lidar_pipeline/visualizations.py
@@ -214,22 +214,19 @@ def _fill_nans(arr):
     """Fill NaN values using nearest-neighbor interpolation.
 
     Returns (filled_array, nan_mask) so the caller can restore NaN after filtering.
+
+    Via transformée de distance (O(n), vectorisé) : les indices du plus proche
+    voisin valides sortent en une passe. NearestNDInterpolator construisait un
+    cKDTree sur TOUS les points valides (25 M à 0,2 m) — plusieurs secondes
+    par dalle trouée, payées au premier accès de SharedDEM.filled.
     """
-    from scipy.interpolate import NearestNDInterpolator
     nan_mask = np.isnan(arr)
     if not np.any(nan_mask):
         return arr, nan_mask
-    valid = ~nan_mask
-    y_coords, x_coords = np.where(valid)
-    if len(y_coords) == 0:
-        return np.zeros_like(arr), nan_mask
-    z_values = arr[valid]
-    interp = NearestNDInterpolator(
-        np.column_stack((y_coords, x_coords)), z_values
-    )
-    y_missing, x_missing = np.where(nan_mask)
+    from scipy.ndimage import distance_transform_edt
+    _, (iy, ix) = distance_transform_edt(nan_mask, return_indices=True)
     filled = arr.copy()
-    filled[y_missing, x_missing] = interp(y_missing, x_missing)
+    filled[nan_mask] = arr[iy[nan_mask], ix[nan_mask]]
     return filled, nan_mask