Automatic image-based skin diagnostics using deep learning
There is shown and described a deep learning based system and method for skin diagnostics as well as testing metrics that show that such a deep learning based system outperforms human experts on the task of apparent skin diagnostics. Also shown and described is a system and method of monitoring a skin treatment regime using a deep learning based system and method for skin diagnostics.
1. A skin diagnostic device comprising:
a memory configured to store and provide a convolutional neural network (CNN) configured to classify pixels of an image to determine a plurality (N) of respective skin sign diagnoses for each of a plurality (N) of respective skin signs wherein the CNN comprises a deep neural network for image classification configured to generate the N respective skin sign diagnoses and wherein the CNN is trained using skin sign data for each of the N respective skin signs; and
at least one processor coupled to the memory and configured to receive the image and process the image using the CNN to generate the N respective skin sign diagnoses,
wherein the CNN comprises:
an encoder phase defined from a pre-trained network for image classification and configured to encode features to a final encoder phase feature net; and
a decoder phase configured to receive the final encoder phase feature net for decoding by a plurality (N) of respective parallel skin sign branches to generate each of the N respective skin sign diagnoses.
2. The skin diagnostic device according to claim 1 , wherein the decoder phase includes a global pooling operation to process the final encoder phase feature net to provide to each of the N respective parallel skin sign branches.
3. The skin diagnostic device according to claim 1 , wherein the CNN is further configured to classify the pixels to determine an ethnicity vector and the CNN is trained using skin sign data for each of the N respective skin signs and a plurality of ethnicities.
4. The skin diagnostics device according to claim 3 , wherein the decoder phase comprises a further parallel branch for ethnicity to generate the ethnicity vector.
5. The skin diagnostic device according to claim 3 :
wherein the CNN is trained using multiple samples in the form (x i , y i ), with x i being the i-th training image and y i being a corresponding vector of ground truth skin sign diagnoses; and
wherein the CNN is trained to minimize a loss function for each respective branch of the N parallel skin sign branches and the further parallel branch for ethnicity.
6. The skin diagnostic device according to claim 5 , wherein the CNN is further trained to minimize a loss function L, comprising a L2 loss function for each of the N respective skin sign branches in a weighted combination with a standard cross-entropy classification loss L ethnicity for the further parallel branch for ethnicity, according to equation (3):
L=L 2+ λL ethnicity (3)
where λ, controls a balance between a score regression and ethnicity classification losses.
7. The skin diagnostic device according to claim 1 , wherein each branch of the N respective parallel skin sign branches comprises in succession: a first fully connected layer followed, a first activation layer, a second fully connected layer, a second activation layer and a final activation layer to output a final value comprising one of the N respective skin sign diagnoses and the ethnicity vector.
8. The skin diagnostic device according to claim 7 , wherein the final activation layer is defined in accordance with a function of equation (1) for an input score x received from the second activation layer:
LeakyClamp
(
x
)
=
{
x
if
x
∈
[
a
,
b
]
α
(
x
-
a
)
+
a
if
x
<
a
α
(
x
-
b
)
+
b
if
x
>
b
(
1
)
where α is a slope, α is a lower bound and b is an upper bound of a respective score range for each the N respective skin sign diagnoses.
9. The skin diagnostic device according to claim 1 , wherein the memory stores a face and landmark detector to pre-process the image and wherein the at least one processor is configured to generate a normalized image from the image using the face and landmark detector and use the normalized image when using the CNN.
10. The skin diagnostic device according to claim 1 , wherein the CNN comprises a pre-trained network for image classification which is adapted to generate the N respective skin sign diagnoses such that:
the fully connected layers of the pre-trained network are removed; and
N respective groups of layers are defined to decode a same feature net for each of the N respective skin sign diagnoses in parallel.
11. The skin diagnostic device according to claim 1 , configured as one of:
a computing device for personal use comprising a mobile device; and
a server providing skin diagnostic services via a communications network.
12. The skin diagnostic device according to claim 1 , wherein the memory stores code which when executed by the at least one processor provides a treatment product selector responsive to at least some of the N skin sign diagnoses to obtain a recommendation for at least one of a product and a treatment plan.
13. The skin diagnostic device according to claim 1 , wherein the memory stores code which when executed by the at least one processor provides an image acquisition function to receive the image.
14. The skin diagnostic device according to claim 1 , wherein the memory stores code which when executed by the at least one processor provides a treatment monitor to monitor treatment for at least one skin sign.
15. The skin diagnostic device according to claim 14 , wherein the at least one processor is configured to at least one of remind, instruct and/or record treatment activities associated with a product application for respective treatment sessions.
16. The skin diagnostic device according to claim 1 , wherein the at least one processor configured to process a second image using the CNN to generate a subsequent skin diagnoses received following a treatment session.
17. The skin diagnostic device according to claim 16 , wherein the memory stores code which when executed by the at least one processor a presentation of comparative results using the subsequent skin diagnoses.
18. A method implemented by a skin diagnostic device comprising:
storing and providing, by a memory of the skin diagnostic device, a convolutional neural network (CNN) configured to classify pixels of an image to determine a plurality (N) of respective skin sign diagnoses for each of a plurality (N) of respective skin signs wherein the CNN comprises a deep neural network for image classification configured to generate the N respective skin sign diagnoses and wherein the CNN is trained using skin sign data for each of the N respective skin signs; and
performing by at least one processor of the skin diagnostic device that is coupled to the memory:
receiving the image; and
processing the image using the CNN to generate the N respective skin sign diagnoses,
wherein the CNN comprises:
an encoder phase defined from a pre-trained network for image classification and configured to encode features to a final encoder phase feature net; and
a decoder phase configured to receive the final encoder phase feature net for decoding by a plurality (N) of respective parallel skin sign branches to generate each of the N respective skin sign diagnoses.
19. A method implemented by at least one processor of a computer comprising:
training a convolutional neural network (CNN) configured to classify pixels of an image to determine a plurality (N) of respective skin sign diagnoses for each of a plurality (N) of respective skin signs wherein the CNN comprises a deep neural network for image classification configured to generate the N respective skin sign diagnoses and wherein the training is performed using skin sign data for each of the N respective skin signs,
wherein the CNN comprises:
an encoder phase defined from a pre-trained network for image classification and configured to encode features to a final encoder phase feature net; and
a decoder phase configured to receive the final encoder phase feature net for decoding by a plurality (N) of respective parallel skin sign branches to generate each of the N respective skin sign diagnoses.