Fully de-duplicate vertices in MarchingCube algorithm

Recently I started using Meshing.isosurface(), with MarchingCubes as method. The docs say that
setting reduceverts=true (default: true) will merge vertices within a voxel to reduce mesh size by around 30%
and with slight performance improvement.
I called both Meshing.isosurface() and skimage.measure.marching_cubes_lewiner(), a Python function from scikit-image,
to extrude a brain within a volumetric data of size (91, 109, 91).
The number of vertices of the corresponding mesh, returned by Meshing function is (345752) almost 4 times greater
than that returned by skimage function (86473).
345752/86473 = 3.99, i.e. it is exaggerated.

Here are the two blocks of code:

using Meshing
using NPZ           
brain_vol = npzread("MNI152.npy") # https://github.com/empet/Datasets/blob/master/MNI152.npy
verts, triangles =  isosurface(brain_vol, MarchingCubes(iso=1.25, reduceverts=true, eps=1e-05))
length(verts),  length(triangles)
(345752, 172932) 
            
#unique vertices:
uni_verts = unique(i -> verts[i], 1:length(verts))
length(uni_verts)
89264 

respectively:

import numpy as np
from skimage import measure            
brain_vol= np.load("MNI152.npy")
verts, triangles = measure.marching_cubes_lewiner(brain_vol, 1.25)[:2] 
len(verts), len(triangles)
(86473, 173070)

The corresponding Jupyter Notebook has the size of 111MB, vs 25MB, when using Python scikit-image.
Otherwise the mesh plot is great in both cases:

Start at Meshing.isosurface with MarchingCubes(iso=1.25, reduceverts=true, eps=1e-05) and reproduce the reported vertex counts using the linked MNI152.npy data. Compare the returned verts with the unique-vertex count and the scikit-image marching_cubes_lewiner result; done means reduceverts fully deduplicates shared vertices without invalidating the triangles.