Table 2 Clustering performance of various features due only to the strongly defined types after digital flattening.

From: A genetic and computational approach to structurally classify neuronal types

 

Hull area

Br. points

Dend length

Med br. len.

Br. angle

Tort.

Asym.

Soma to str.

Typical radius

Med z -pos

Σ

Our method*

Silhouette

0.15

−0.13

−0.11

−0.08

−0.09

−0.07

0.01

0.01

0.06

0.77

0.75

0.77

Structural confusions

11

9

14

10

15

13

15

10

10

2

1

0

Genetic confusions

11

9

13

10

14

13

15

11

10

1

0

0

Total confusions

22

18

27

20

29

26

30

21

20

3

1

0

  1. Asym.: asymmetry; Br. angle, average branching angle; Br. points, branch point count; Dend. length, dendritic length; Med. br. len., median branch length; Med. z-pos.: median z-position; Soma to str., soma-to-stratification depth distance; Tort., tortuosity.
  2. Bold columns use registered depth information.
  3. *The filtering parameters in our method were optimized over the entire genetically defined data set for consistency. Filtering parameters optimized over only the strongly defined cell types would result in an even higher silhouette value. Σ denotes an optimal combination of the median z-position, typical radius, and the dendritic length. Higher is better for silhouette scores, 1 being the theoretical maximum. Lower is better for confusion scores, 0 denoting no confusions (see Methods for feature and performance metric definitions).
Back to article page