1
00:00:01,320 --> 00:00:05,960
In this second video of our two-part
Deep-Learning Technology video series,

2
00:00:05,960 --> 00:00:09,480
we’ll show you how to use the training data
we created in part 1

3
00:00:09,480 --> 00:00:13,800
to setup your software for label-free nucleus detection.

4
00:00:13,800 --> 00:00:16,320
To start, go to the deep learning layout tab

5
00:00:16,320 --> 00:00:23,720
and click new training in the deep learning tool window.

6
00:00:23,720 --> 00:00:25,640
Change the name of your new training

7
00:00:25,640 --> 00:00:30,480
and select the data set you created
while watching the first video.

8
00:00:35,520 --> 00:00:40,120
Click OK on the dialog box and wait a few seconds.

9
00:00:50,080 --> 00:00:52,960
Then, choose the input channel.

10
00:00:52,960 --> 00:00:58,280
For label-free nucleus detection,
you’ll select the transmitted light channel.

11
00:00:58,280 --> 00:01:02,120
However, since we’re going to use
the training data to detect nuclei

12
00:01:02,120 --> 00:01:07,000
from phase contrast images,
select PH instead of transmitted light.

13
00:01:07,000 --> 00:01:12,520
The training label class is nuclei only,
and we defined it in the previous video.

14
00:01:12,520 --> 00:01:15,960
Now we can select the training configuration.

15
00:01:15,960 --> 00:01:19,360
In many cases, the standard network is OK.

16
00:01:19,360 --> 00:01:22,800
The large network is appropriate for more complicated structures,

17
00:01:22,800 --> 00:01:24,600
but it takes more time.

18
00:01:24,600 --> 00:01:31,560
Let’s keep the default iteration.
If necessary, we can increase it later.

19
00:01:31,560 --> 00:01:37,240
We used an NVIDIA Quadro P2000,
and the process took about 50 minutes.

20
00:01:37,240 --> 00:01:40,440
There are several ways to check accuracy.

21
00:01:40,440 --> 00:01:42,840
The similarity between the training layer

22
00:01:42,840 --> 00:01:45,960
and the layers where the neural network
identified the nuclei,

23
00:01:45,960 --> 00:01:49,240
called inferred layers, is shown by default.

24
00:01:49,240 --> 00:01:53,160
The minimum value is 0, and the max is 1.

25
00:01:53,160 --> 00:01:57,680
Note that the similarity increases during the training process.

26
00:01:57,680 --> 00:02:01,600
When each iteration finishes, a check point is created.

27
00:02:01,600 --> 00:02:03,760
You can check the quality of the neural network

28
00:02:03,760 --> 00:02:06,560
by validating the data at each check point.

29
00:02:06,560 --> 00:02:10,120
Validation data are selected from the training data set.

30
00:02:10,120 --> 00:02:15,040
The red areas are the inferred nuclei from
the validation data’s phase contrast channel.

31
00:02:15,040 --> 00:02:19,560
The red circle is the training label,
showing the correct nuclei area.

32
00:02:19,560 --> 00:02:22,120
In this example, the data looks OK,

33
00:02:22,120 --> 00:02:24,320
but there are a few missed detections.

34
00:02:24,320 --> 00:02:27,880
You can save the neural network at
a check point and end the training,

35
00:02:27,880 --> 00:02:29,960
but let’s continue.

36
00:02:29,960 --> 00:02:32,880
In this example, the training process is finished

37
00:02:32,880 --> 00:02:35,800
after five check points have been created.

38
00:02:35,800 --> 00:02:37,880
Now that the training is complete,

39
00:02:37,880 --> 00:02:40,360
we can see that the detection is very accurate

40
00:02:40,360 --> 00:02:45,080
and the errors we saw at the first check point are gone.

41
00:02:45,080 --> 00:02:47,600
Save the neural network at checkpoint five

42
00:02:47,600 --> 00:02:55,280
so that we can use it for other images.

43
00:02:55,280 --> 00:02:58,280
The neural network training is complete.

44
00:02:58,280 --> 00:03:00,680
Go to the Count and Measure layout tab

45
00:03:00,680 --> 00:03:04,240
to apply the neural network to other images.

46
00:03:04,240 --> 00:03:09,440
The images we’re going to use in this example
are available in the installer guide.

47
00:03:09,440 --> 00:03:16,320
Open two of the images.

48
00:03:16,320 --> 00:03:21,360
Apply the neural network to the
phase contrast channel to infer nuclei.

49
00:03:21,360 --> 00:03:26,120
Then apply the Count and Measure function
to the inferred nuclei to count them.

50
00:03:26,120 --> 00:03:30,200
Record this process as a one-click macro.

51
00:03:30,200 --> 00:03:35,280
Disable toggle batch macro if needed
and click create macro.

52
00:03:35,280 --> 00:03:47,920
Define the name of the macro and start recording.

53
00:03:47,920 --> 00:03:51,080
From the menu bar, select neural network processing

54
00:03:51,080 --> 00:03:57,160
and then select the neural network.

55
00:03:57,160 --> 00:03:59,360
The input channel must be PH

56
00:03:59,360 --> 00:04:03,920
since this neural network was trained
using the phase contrast channel.

57
00:04:03,920 --> 00:04:15,280
We recommend unchecking the create new document
as output box for macro recording.

58
00:04:15,280 --> 00:04:18,520
The resulting inference layer identifying the nuclei

59
00:04:18,520 --> 00:04:21,040
is shown as a probability layer.

60
00:04:21,040 --> 00:04:25,640
The red area shows the inferred nuclei from the PH channel.

61
00:04:25,640 --> 00:04:29,160
The gray value isn’t the intensity of the DAPI fluorescence,

62
00:04:29,160 --> 00:04:32,840
but the probability of an object being a nucleus.

63
00:04:32,840 --> 00:04:35,200
While this image includes the DAPI channel,

64
00:04:35,200 --> 00:04:37,400
it’s only to verify the data.

65
00:04:37,400 --> 00:04:41,000
It’s not used for the inference process.

66
00:04:41,000 --> 00:04:45,840
You can conduct additional analysis
using the Count and Measure module.

67
00:04:45,840 --> 00:04:49,200
Click to activate the probability layer.

68
00:04:49,200 --> 00:04:53,360
A probability threshold will be applied to the nuclei.

69
00:04:53,360 --> 00:04:56,880
Set the threshold so that two close nuclei
are properly split

70
00:04:56,880 --> 00:05:05,680
and so that all nuclei are properly detected.

71
00:05:05,680 --> 00:05:09,720
The measurement results for each object are shown in the chart.

72
00:05:09,720 --> 00:05:13,520
At this point, you can stop the macro recording.

73
00:05:13,520 --> 00:05:27,200
Confirm that the detected object layer has been created.

74
00:05:27,200 --> 00:05:33,520
This analysis process can be applied to your
other images by clicking run macro.

75
00:05:49,600 --> 00:05:53,320
You can also go back and apply a threshold to the DAPI channel

76
00:05:53,320 --> 00:05:56,240
to compare it with the probability channel.

77
00:05:56,240 --> 00:06:06,600
Select the DAPI channel and change the color for easy checking.

78
00:06:06,600 --> 00:06:10,680
As you can see, it’s impossible to separate two close nuclei

79
00:06:10,680 --> 00:06:13,680
no matter how we adjust threshold.

80
00:06:13,680 --> 00:06:15,920
If we separate the upper two objects,

81
00:06:15,920 --> 00:06:19,680
many other nuclei will not be properly detected.

82
00:06:19,680 --> 00:06:23,440
This demonstrates that the neural network’s
label-free nucleus detection

83
00:06:23,440 --> 00:06:28,760
using the phase contrast image works better
than the traditional DAPI staining.

84
00:06:28,760 --> 00:06:32,400
In addition to being able to identify nuclei without staining,

85
00:06:32,400 --> 00:06:35,840
the deep-learning technology helps keep your cells healthier

86
00:06:35,840 --> 00:06:40,440
by eliminating the phototoxicity
caused by fluorescence excitation.

87
00:06:40,440 --> 00:06:45,160
If you’d like more information about
cellSens software’s deep-learning technology,

88
00:06:45,160 --> 00:06:48,240
visit
olympus-lifescience.com.

89
00:06:48,240 --> 00:06:49,960
Thanks for watching.