Conjoined twin network for treatment and analysis
An apparatus for treating an abnormality includes a processor and a non-transitory computer readable medium that includes a first convolutional neural network (CNN) having weights and a second CNN in parallel with the first CNN and sharing the weights, the second CNN being joined to the first neural network by a distance function. The medium includes instructions that when executed by the processor implements a method. The method includes receiving a first image dataset of an area of interest and processing the first image dataset using the first CNN and receiving a second image dataset of the area of interest obtained prior to the first image dataset and processing the second image dataset using the second CNN. The method also includes identifying the abnormality using an output of the distance function wherein the identifying influences treatment of the abnormality.
1. An apparatus for treating an abnormality, the apparatus comprising:
a processor;
a non-transitory computer readable medium comprising:
a first convolutional neural network (CNN) having weights;
a second convolutional neural network in parallel with the first CNN and sharing the weights of the first CNN, the second CNN being joined to the first neural network by a distance function;
instructions that when executed by the processor implement a method comprising:
receiving a first image dataset of an area of interest and processing the first image dataset using the first CNN;
receiving a second image dataset of the area of interest obtained prior to the first image dataset and processing the second image dataset using the second CNN;
identifying the abnormality based on an output of the distance function; and
outputting an indication of the abnormality, wherein the indication influences administering or adjusting treatment of the abnormality;
wherein the first CNN comprises a first series of neural network layers that provide a current feature vector f C and the second CNN comprises a second series of neural network layers that provide a prior feature vector f P , f C and f P being input to a first distance function that provides output d 1 and a second distance function that provides output d 2 , d 1 and d 2 being input to a sigmoid function that identifies the abnormality.
2. The apparatus according to claim 1 , wherein the abnormality is of a patient and the treatment comprises at least one of surgery, chemotherapy, hormonal therapy, immunotherapy, or radiation therapy.
3. The apparatus according to claim 1 , wherein the abnormality is of a structural element and the treatment comprises at least one of repair or replacement of the structural element.
4. The apparatus according to claim 1 , wherein the first CNN and the second CNN are trained using a plurality annotated training image datasets.
5. The apparatus according to claim 4 , where the first CNN and the second CNN are trained using a loss function comprising a linear combination of an entropy term, an L1 norm term, and an L2 norm term.
6. The apparatus according to claim 1 , wherein the method further comprises identifying a location of the abnormality.
7. The apparatus according to claim 6 , wherein the current feature vector f C and the prior feature vector f P are provided to a feature correlation module (FCM) comprising a matrix subtraction function and a SiLU activation function that outputs a distance value.
8. The apparatus according to claim 7 , further comprising an attention suppress gate module (ASGM) coupled to an output of the FCM, the ASGM comprising a Hadamard product function and a sigmoid activation function coupled to an output of the Hadamard product function.
9. The apparatus according to claim 8 , further comprising a breast abnormal module (BAM) comprising a convolution layer configured to blend extracted feature vectors, a sigmoid function coupled to an output of the convolution layer, and a threshold value such that an output of the sigmoid function being greater than or equal to the threshold value provides indication of the location of the abnormality.
10. The apparatus according to claim 9 , further comprising a loss function module, the loss function module comprising a similarity index measurement (SSIM) reconstruction loss, a binary cross function (BE), and an L2 norm function.
11. The apparatus according to claim 6 , wherein the first CNN and the second CNN are trained using a plurality unannotated training image datasets.
12. A non-transitory computer readable medium for treating an abnormality comprising:
a first convolutional neural network (CNN) having weights;
a second convolutional neural network in parallel with the first CNN and sharing the weights of the first CNN, the second CNN being joined to the first CNN by a distance function;
instructions that when executed by the processor implements a method comprising:
receiving a first image dataset of an area of interest and processing the first image dataset using the first CNN;
receiving a second image dataset of the area of interest obtained prior to the first image dataset and processing the second image dataset using the second CNN;
identifying the abnormality based on an output of the distance function; and
outputting an indication of the abnormality, wherein the indication influences administering or adjusting treatment of the abnormality;
wherein the first CNN comprises a first series of neural network layers that provide a current feature vector f C and the second CNN comprises a second series of neural network layers that provide a prior feature vector f P , f C and f P being input to a first distance function that provides output d 1 and a second distance function that provides output d 2 , d 1 and d 2 being input to a sigmoid function that identifies the abnormality.
13. The non-transitory computer readable medium according to claim 12 , wherein the method further comprises identifying a location of the abnormality.
14. The non-transitory computer readable medium according to claim 13 , wherein the current feature vector f C and the prior feature vector f P are provided to a feature correlation module (FCM) comprising a matrix subtraction function and a SiLU activation function that outputs a distance value.
15. The non-transitory computer readable medium according to claim 14 , further comprising an attention suppress gate module (ASGM) coupled to an output of the FCM, the ASGM comprising a Hadamard product function and a sigmoid activation function coupled to an output of the Hadamard product function.
16. The non-transitory computer readable medium according to claim 15 , further comprising a breast abnormal module (BAM) comprising a convolution layer configured to blend extracted feature vectors, a sigmoid function coupled to an output of the convolution layer, and a threshold value such that an output of the sigmoid function being greater than or equal to the threshold value provides indication of the location of the abnormality.
17. A method for treating an abnormality, the method comprising:
receiving a first image dataset of an area of interest and processing the first image dataset using a first convolutional neural network (CNN), the first CNN having weights;
receiving a second image dataset of the area of interest obtained prior to the first image dataset and processing the second image dataset using a second convolutional neural network, the second CNN being in parallel with the first CNN and sharing the weights of the first CNN, the second CNN being joined to the first neural network by a distance function;
identifying the abnormality based on an output of the distance function; and
influencing application or adjustment of treatment of the abnormality based on identification of the abnormality;
wherein the first CNN comprises a first series of neural network layers that provide a current feature vector f C and the second CNN comprises a second series of neural network layers that provide a prior feature vector f P , the method further comprising:
inputting f C and f P into a first distance function that provides output d 1 ;
inputting f C and f P into a second distance function that provides output d 2 ; and
inputting d 1 and d 2 into a sigmoid function that identifies the abnormality.
18. The method according to claim 17 , further comprising identifying a location of the abnormality.
19. The method according to claim 18 , further comprising inputting f C and f P into a feature correlation module (FCM) comprising a matrix subtraction function and a SiLU activation function that outputs a distance value.
20. The method according to claim 17 , wherein the abnormality is of a patient and the treatment comprises at least one of surgery, chemotherapy, hormonal therapy, immunotherapy, or radiation therapy.
21. The method according to claim 17 , wherein the abnormality is of a structural element and the treatment comprises at least one of repair or replacement of the structural element.