Systems for identifying cells and uses thereof
Disclosed herein are a system for identifying cells on a microscopic image. According to some embodiments, the system comprises a non-transitory processor-readable medium, and a processor communicably configured to receive the microscopic image, and process the received microscopic image with a convolutional neural network (CNN) model having a modified U-Net architecture. Also disclosed herein are methods for identifying a spatial pattern of human induced pluripotent stem cells (hiPSCs) by using the present system.
1. A system for identifying cells on a microscopic image, comprising,
a non-transitory processor-readable medium that stores processor-readable instructions; and
a processor communicably coupled to the non-transitory processor-readable medium and configured to,
receive the microscopic image; and
process the received microscopic image with a convolutional neural network (CNN) model, wherein the CNN model has a U-Net architecture comprising,
a plurality of down-sampling sets for extracting the features of the microscopic image thereby generating feature maps; and
a plurality of up-sampling sets respectively for generating a segmented image from the feature maps; wherein
each down-sampling set comprises at least one convolution layer and at least one pooling layer preceded by the convolution layer; and
the extraction is carried out by the plurality of down-sampling sets on a set-by-set basis by using equation (1),
(
x
,
y
,
C
)
=
(
h
S
i
-
1
,
w
S
i
-
1
,
2
F
+
i
-
1
)
,
(
1
)
where x and y are respectively the pixels of the height and width of an input image processed by each down-sampling set, C is the channel number of the input image, h and w are respectively the pixels of the height and width of the microscopic image, S is the stride of the pooling layer, i is the number of each down-sampling set in the plurality of down-sampling sets, and F is a constant for deciding the channel number of the input image.
2. The system of claim 1 , wherein the pooling layer is a max-pooling layer.
3. The system of claim 2 , wherein the S and F are respectively set to be 2 and 5.
4. The system of claim 1 , wherein the CNN model is established by a CNN training method comprising,
(a) inputting a plurality of training images to the CNN model to generate a plurality of predicted images;
(b) generating a plurality of ground truth images respectively corresponding to the plurality of training images; and
(c) determining a loss function for the CNN model based on the predicted images of step (a) and the ground truth images of step (b) by using equations (2)-(5),
MPL
(
X
,
Y
)
=
α
×
W
(
X
,
Y
)
+
β
×
D
(
X
,
Y
)
,
(
2
)
W
(
X
,
Y
)
=
1
N
∑
j
=
1
N
E
(
X
j
,
Y
j
)
×
γ
j
,
(
3
)
E
(
X
j
,
Y
j
)
=
-
Y
j
×
log
X
j
,
(
4
)
D
(
X
,
Y
)
=
1
-
2
❘
"\[LeftBracketingBar]"
X
⋂
Y
❘
"\[RightBracketingBar]"
/
(
❘
"\[LeftBracketingBar]"
X
❘
"\[RightBracketingBar]"
+
❘
"\[LeftBracketingBar]"
❘
"\[LeftBracketingBar]"
Y
❘
"\[RightBracketingBar]"
)
,
(
5
)
where W(X,Y) is a weight loss function, D(X,Y) is a dice loss function, α is the weight of W(X,Y), β is the weight of D(X,Y), N is the number of the training images, X is the pixels of each predicted image, Y is the pixels of each ground truth image, and γ is the weight of j category.
5. The system of claim 4 , wherein the α and β are respectively set to be 1.1 and 0.9.
6. A non-transitory processor-readable medium storing processor-readable instructions, wherein the processor-readable instructions, when executed at a processor of a computing device, are capable of,
receiving a microscopic image; and
processing the received microscopic image with a convolutional neural network (CNN) model, wherein the CNN model has a U-Net architecture comprising,
a plurality of down-sampling sets for extracting the features of the microscopic image thereby generating feature maps; and
a plurality of up-sampling sets for generating a segmented image from the feature maps; wherein
each down-sampling set comprises at least one convolution layer and at least one pooling layer preceded by the convolution layer; and
the extraction is carried out by the plurality of down-sampling sets on a set-by-set basis by using equation (1),
(
x
,
y
,
C
)
=
(
h
S
i
-
1
,
w
S
i
-
1
,
2
F
+
i
-
1
)
,
(
1
)
where x and y are respectively the pixels of the height and width of an input image processed by each down-sampling set, C is the channel number of the input image, h and w are respectively the pixels of the height and width of the microscopic image, S is the stride of the pooling layer, i is the number of each down-sampling set in the plurality of down-sampling sets, and F is a constant for deciding the channel number of the input image.
7. The non-transitory computer-readable medium of claim 6 , wherein the pooling layer is a max-pooling layer.
8. The non-transitory computer-readable medium of claim 7 , wherein the S and F are respectively set to be 2 and 5.
9. The non-transitory computer-readable medium of claim 6 , wherein the CNN model is established by a CNN training method comprising,
(a) inputting a plurality of training images to the CNN model to generate a plurality of predicted images;
(b) generating a plurality of ground truth images respectively corresponding to the plurality of training images; and
(c) determining a loss function for the CNN model based on the predicted images of step (a) and the ground truth images of step (b) by using equations (2)-(5),
MPL
(
X
,
Y
)
=
α
×
W
(
X
,
Y
)
+
β
×
D
(
X
,
Y
)
,
(
2
)
W
(
X
,
Y
)
=
1
N
∑
j
=
1
N
E
(
X
j
,
Y
j
)
×
γ
j
,
(
3
)
E
(
X
j
,
Y
j
)
=
-
Y
j
×
log
X
j
,
(
4
)
D
(
X
,
Y
)
=
1
-
2
❘
"\[LeftBracketingBar]"
X
⋂
Y
❘
"\[RightBracketingBar]"
/
(
❘
"\[LeftBracketingBar]"
X
❘
"\[RightBracketingBar]"
+
❘
"\[LeftBracketingBar]"
❘
"\[LeftBracketingBar]"
Y
❘
"\[RightBracketingBar]"
)
,
(
5
)
where W(X,Y) is a weight loss function, D(X,Y) is a dice loss function, α is the weight of W(X,Y), β is the weight of D(X,Y), N is the number of the training images, X is the pixels of each predicted image, Y is the pixels of each ground truth image, and γ is the weight of j category.
10. The system of claim 9 , wherein the α and β are respectively set to be 1.1 and 0.9.
11. A method for identifying a spatial pattern of human induced pluripotent stem cells (hiPSCs), comprising,
(a) respectively labeling the nuclei and one of the germ layers of the hiPSCs with a first and a second molecules;
(b) respectively obtaining a first microscopic image of the nucleus-labeled hiPSCs and a second microscopic image of the germ layer-labeled hiPSCs;
(c) producing a segmented image by processing the first microscopic image with the system of claim 1 ;
(d) superimposing the segmented image produced by step (c) on the second microscopic image of step (b); and
(e) identifying the spatial pattern of the hiPSCs based on the superimposed image of step (d).
12. The method of claim 11 , wherein the first molecule is a nuclear dye, and the second molecule is an antibody against one of the germ layers of the hiPSCs.
13. The method of claim 12 , wherein the antibody is an anti-ectoderm, anti-mesoderm or anti-endoderm antibody.