Retina vessel measurement
Disclosed is a method for training a neural network to quantify the vessel calibre of retina fundus images. The method involves receiving a plurality of fundus images; pre-processing the fundus images to normalise images features of the fundus images; and training a multi-layer neural network, the neural network comprising of a convolutional unit, multiple dense blocks alternating with transition units for down-sampling image features determined by the neural network, and a fully-connected unit, wherein each dense block comprises a series of cAdd units packed with multiple convolutions, and each transition layer comprises a convolution with pooling.
1. A method for training a neural network for automated retina vessel measurement, comprising:
receiving a plurality of fundus images;
pre-processing the fundus images to normalise images features of the fundus images; and
training a multi-layer neural network on the pre-processed fundus images, the neural network comprising convolutional unit, multiple dense blocks alternating with transition layers or transition units for down-sampling image features determined by the neural network, and a fully-connected unit, wherein each dense block comprises a series of cAdd units packed with multiple convolutions, and each transition layer or transition unit comprises a convolution with pooling.
2. The method of claim 1 , further comprising grouping the input channels of each cAdd unit into non-overlapping groups and adding the outputs of the cAdd unit to one of the non-overlapping groups, thus forming the inputs to a next cAdd unit in the series, and for successive ones of said cAdd units, the output of a previous cAdd unit in the series is added to a different one of said non-overlapping groups.
3. The method of claim 1 , further comprising:
automatically detecting a centre of an optic disc in each fundus image; and
cropping the respective image to a region of predetermined dimensions centred on the optic disc centre.
4. The method of claim 1 , wherein pre-processing the fundus images comprises applying global contrast normalisation to each fundus image.
5. The method of claim 3 , wherein pre-processing the fundus images further comprises median filtering using a kernel of predetermined size.
6. The method of 1 , wherein there are five dense blocks in the multiple dense blocks.
7. The method of claim 1 , wherein each dense block comprises a series of cAdd units packed with two types of convolutions.
8. The method of claim 7 , wherein the two types of convolutions comprise a 1×1 convolution and a 3×3 convolution.
9. The method of claim 1 , wherein the convolution of each transition layer or transition unit is a 1×1 convolution.
10. A method of quantifying vessel calibre of a retina fundus image, comprising:
receiving a retina fundus image; and
applying a neural network trained according to the method of claim 1 to the retina fundus image.
11. A computer system for training a neural network to generate retina vessel measurements, comprising:
a memory; and
at least one processor, the memory storing a multi-layer neural network and instructions that, when executed by the at least one processor cause the at least one processor to:
receive a plurality of fundus images;
pre-process the fundus images to normalise images features of the fundus images; and
train the neural network on the pre-processed fundus images, the neural network comprising a convolutional unit, multiple dense blocks alternating with transition units or transition layers for down-sampling image features determined by the neural network, and a fully-connected unit, wherein each dense block comprises a series of cAdd units packed with multiple convolutions, and each transition unit or transition layer comprises a convolution with pooling.
12. The computer system of claim 11 , wherein the instructions further cause the processor to group the input channels of each cAdd unit into non-overlapping groups and adding the outputs of the cAdd unit to one of the non-overlapping groups, thus forming the inputs to a next cAdd unit in the series, and for successive ones of said cAdd units, the output of a previous cAdd unit in the series is added to a different one of said non-overlapping groups.
13. The computer system of claim 11 , wherein the instructions further cause the processor to:
automatically detect a centre of an optic disc in each fundus image; and
crop the respective image to a region of predetermined dimensions centred on the optic disc centre.
14. The computer system of claim 11 , wherein the instructions cause the processor to pre-process the fundus images by applying global contrast normalisation to each fundus image.
15. The computer system of claim 11 , wherein there are five dense blocks in the multiple dense blocks.
16. The computer system of claim 11 , wherein each dense block comprises a series of cAdd units packed with two types of convolutions.
17. The computer system of claim 16 , wherein the two types of convolutions comprise a 1×1 convolution and a 3×3 convolution.
18. The computer system of claim 11 , wherein the convolution of each transition unit or transition layer is a 1×1 convolution.
19. The computer system of claim 11 , wherein the neural network is trained on the pre-processed fundus images to quantify a vessel calibre of a retina fundus image.