Skin lesion segmentation using deep convolution networks guided by local unsupervised learning

1. A non-transitory computer readable medium comprising computer executable instructions which when executed by a computer cause the computer to perform the method of:

obtaining a dermoscopic image;

running a convolutional neural network image classifier on the dermoscopic image to obtain pixelwise prediction scores for presence or absence of a lesion;

segmenting the dermoscopic image into a plurality of super-pixels;

computing for each super-pixel a prediction score as the average of the pixelwise prediction scores for all pixels within that super-pixel;

computing a mean prediction score across the plurality of super-pixels;

constructing an indicator vector Y in which a confidence indicator of “1” is assigned to each super-pixel with a prediction score equal or greater than the mean prediction score, and a confidence indicator of “0” is assigned to each super-pixel with a prediction score less than the mean prediction score;

constructing a super-pixel graph G=(V,E,W) wherein

w

ij

=

exp

⁡

(

-



x

i

-

x

j



2

σ

)

 and d i =Σ i=1 N W ij , wherein V is a matrix of vertices representing the super-pixels, E is a matrix of edges representing similarities among the super-pixels, W is a matrix of weights of edges representing degrees of similarity among the super-pixels, w ij are elements of W, x i , and x j are individual super-pixels, σ is a constant that scales the strength of each element of W, and d i is a degree of super-pixel x i ;

computing a confidence score function F according to {circumflex over (F)}=arg min(F T LF +μ∥F−Y∥ 2 ), wherein L is the Laplacian of F and {circumflex over (F)}=(D−λW) -1 Y, wherein D is a matrix of d i , λ=1/1+μ and μ=0.01; and

integrating the confidence score function F with the pixelwise prediction scores to produce a final segmentation of the dermoscopic image into lesion and background areas.

2. The product of claim 1 wherein the confidence score function F is integrated with the pixelwise prediction scores as the vector of image pixels confidence scores.

3. The product of claim 1 wherein the final segmentation has a lesion area where at least two adjacent pixels have prediction scores >0.5 and has a background area where at least two adjacent pixels have prediction scores <=0.5.

4. The product of claim 1 further comprising augmenting the dermoscopic image by rotating a lesion area of the dermoscopic image at least once by .pi./4 radians.

5. The product of claim 4 further comprising augmenting the dermoscopic image by rotating the lesion area at least four times by .pi./4 radians.

6. An apparatus comprising:

a memory; and

at least one processor, coupled to said memory, and operative to implement a method comprising:

obtaining a dermoscopic image;

running a convolutional neural network image classifier on the dermoscopic image to obtain pixelwise prediction scores for presence or absence of a lesion;

segmenting the dermoscopic image into a plurality of super-pixels;

computing for each super-pixel a prediction score as the average of the pixelwise prediction scores for all pixels within that super-pixel;

computing a mean prediction score across the plurality of super-pixels;

constructing an indicator vector Y in which a confidence indicator of “1” is assigned to each super-pixel with a prediction score equal or greater than the mean prediction score, and a confidence indicator of “0” is assigned to each super-pixel with a prediction score less than the mean prediction score;

constructing a super-pixel graph G=(V,E,W) wherein

w

ij

=

exp

⁡

(

-



x

i

-

x

j



2

σ

)

 and d i =Σ i=1 N W ij ,wherein V is a matrix of vertices representing the super-pixels, E is a matrix of edges representing similarities among the super-pixels, W is a matrix of weights of edges representing degrees of similarity among the super-pixels, w ij are elements of W, x i and x j are individual super-pixels, σ is a constant that scales the strength of each element of W, and d i is a degree of super-pixel x i ;

computing a confidence score function F according to {circumflex over (F)}=arg min(F T LF +μ∥F−Y∥ 2 ), wherein L is the Laplacian of F and {circumflex over (F)}=(D−λW) -1 Y, wherein D is a matrix of d i , wherein λ=1/1+μ and μ=0.01; and

integrating the confidence score function F with the pixelwise prediction scores to produce a final segmentation of the dermoscopic image into lesion and background areas.

7. The apparatus of claim 6 wherein the confidence score function F is integrated with the pixelwise prediction scores as the vector of image pixels confidence scores.

8. The apparatus of claim 6 wherein the final segmentation has a lesion area where at least two adjacent pixels have prediction scores >0.5 and has a background area where at least two adjacent pixels have prediction scores <=0.5.

9. The apparatus of claim 6 further comprising augmenting the dermoscopic image by rotating a lesion area of the dermoscopic image at least once by .pi./4 radians.

10. The apparatus of claim 9 further comprising augmenting the dermoscopic image by rotating the lesion area at least four times by .pi./4 radians.

11. The apparatus of claim 9 wherein the convolutional neural network is based on the Oxford Visual Geometry Group network.