This method tested for morphological differences between
anomalous and normal.
Probes were imaged and images were segmented into regular
128x128 pixel segments. Artificial Neural Networks (ANN) were
trained to distinguish segments from condition anomalous
from segments from condition
normal.
To ensure what was learned was not to distinguish specific
instances of CELL LINE and their
distribution between the classes, we trained one ANN per
CELL LINE. We excluded all image
segments from that CELL LINE from
the training, and then used each ANN to compute scores for all the
image segments excluded in their respective training.
Experimental Conditions
The data was separated into conditions anomalous and
normal as follows:
MEASUREMENT
POSITION
DATE
VESSEL
TYPE
VESSEL
MEASURED ON
CELL
LINE
CLASS_NAME
VALID
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
Identifying Dimensions
The classes are completely defined by the values in the metadata
dimensions CLASS_NAME or CELL
LINE. Each of these dimensions may explain any differences
VAIDR may have found in the image data unless it
was selected as the grouping criterion (see above).
For each cut-out, a score was computed, which reflects the
probability to belong to one of the two classes. A score below
0.5 means a classification into the first class, above
0.5 means classification into the second class. For each
well, the score histogram is represented as a violin plot.
Additionally, scores were aggregated for all wells by selecting the
median score of all cut-outs. Median well scores are shown as
circular markers. The wells can be grouped by selectable conditions
and the circular markers can be turned into pie-charts, reflecting
selectable conditions.
Classes
anomalous, normal
|
Group by
|
|
Pie Charts
Binary Classification
Scores▼
To test whether it was possible to distinguish wells assigned to
anomalous from wells assigned
to class normal, aggregate
scores for each well were computed by selecting the median scores
for each of the classes across all cut-outs belonging to that well.
A classification score was computed for each well by dividing the
aggregate score for anomalous
by the sum of the aggregate scores for anomalous and
normal. Classification scores were used
for a Box Plot and a ROC Curve Plot.
Classes
anomalous, normal
|
Group by
|
|
Pie Charts
The mean score for wells from anomalous was significantly
smaller than
that for normal (p = 0.00000
< 0.001)
Statistics on Population Score Means
As a simple means to determine whether our ANNs were able to
learn to distinguish the conditions reliably and robustly, we
computed the mean image segment scores per population (well/flask)
and compared them to the threshold of 0.5.
Populations with a mean score ≥ 0.5 and < 0.5 were counted
classified as anomalous and
normal, respectively. From
this classification we were able to compute prediction
accuracy.
Prediction accuracy was 0.94. The base rate that would have been
achieved by always selecting the majority class was 0.63.
Statistically, this result is highly significant (p ≪
0.001).
The best possible p-value given this number of conditions (72)
and base rate would have been p = 0.00000.
Interpretation
We can conclude that the ANNs we trained learned to distinguish
between image segments from populations from anomalous and
normal. Barring alternative effects which
may have systematically affected cell morphology, this result
supports the hypothesis that cells from the two conditions do have
morphological differences.
