Medical data processing device performing processing using a complex-number neural network
A medical data processing device according to an embodiment includes a processing circuit. The processing circuit applies a linear calculation of a complex number coefficient and a non-linear activation not dependent on an argument of complex numbers, to medical data having a complex value.
1 . A medical data processing device, comprising:
a processing circuit configured to apply, to medical data having a complex value, (1) a linear calculation that uses a complex number coefficient, and (2) a non-linear activation function of a neural network, wherein the activation function is not dependent on an argument of a complex number input into the neural network, the medical data being obtained by processing a signal received from a medical image diagnosis apparatus,
wherein the processing circuit includes the neural network having the non-linear activation function, the neural network outputting output medical data in response to the complex-valued medical data being input to the neural network, and
wherein the linear calculation performed by the processing circuit includes a process in which a bias term is fixed to zero.
2 . The medical data processing device according to claim 1 , wherein the non-linear activation function is a function including a real variable function related to an absolute value of the input complex number.
3 . The medical data processing device according to claim 2 , wherein the non-linear activation function is a function obtained by multiplying the real variable function by the input complex number.
4 . The medical data processing device according to claim 2 , wherein the real variable function included in the non-linear activation function is one of a soft-shrink function, a tanh-shrink function, a power function, and a ReLU function.
5 . The medical data processing device according to claim 1 , wherein the non-linear activation function is a function of which an output value is zero when an input value thereto is zero.
6 . The medical data processing device according to claim 1 , wherein the processing circuit is further configured to apply the non-linear activation function based on a plurality of mutually-different function systems, to the medical data.
7 . The medical data processing device according to claim 1 , wherein the processing circuit is further configured to apply the non-linear activation further to the medical data, by using a plurality of functions which that are based on a mutually same function system and of which function parameters are mutually different.
8 . The medical data processing device according to claim 1 , wherein
the processing circuit is further configured to optimize a function parameter related to the non-linear activation function.
9 . The medical data processing device according to claim 1 , wherein the processing circuit is further configured to apply another activation function in which an output changes depending on the argument of the input complex number.
10 . The medical data processing device according to claim 9 , wherein the processing circuit is further configured to sequentially apply the non-linear activation function that is not dependent on the argument of the input complex number, and the another activation function in which the output changes depending on the argument of the input complex number.
11 . The medical data processing device according to claim 9 , wherein the processing circuit is further configured to selectively apply the non-linear activation function that is not dependent on the argument of the input complex number and the another activation function in which the output changes depending on the argument of the input complex number.
12 . The medical data processing device according to claim 1 , wherein the linear calculation performed by the processing circuit includes a complex convolution.
13 . The medical data processing device according to claim 1 , wherein the medical data is magnetic resonance data.
14 . The medical data processing device according to claim 1 , wherein the medical data is ultrasound data.
15 . The medical data processing device of claim 1 , wherein the medical data is data obtained from scanning a subject using the medical image diagnosis apparatus,
wherein the processing circuit is further configured to cause a display to display the output medical image that is output from the neural network and corresponding to the subject.
16 . The medical data processing device of claim 1 , wherein the medical image diagnosis apparatus is one of an ultrasound scanner, a Magnetic Resonance Imaging (MRI) scanner, and a Position Emission Tomography (PET) scanner.
17 . The medical data processing device of claim 1 , wherein the neural network is a program incorporated in the processing circuit, which is incorporated into one of an ultrasound scanner, a Magnetic Resonance Imaging (MRI) scanner, and a Position Emission Tomography (PET) scanner.
18 . A data processing device, comprising:
a processing circuit configured to apply, to magnetic resonance data having a complex value, (1) a linear calculation that uses a complex number coefficient, and (2) a non-linear activation function of a neural network, wherein the activation function is not dependent on an argument of a complex number input into the neural network, the magnetic resonance data being obtained by processing a signal received from a magnetic resonance imaging apparatus,
wherein the processing circuit includes the neural network having the non-linear activation function, the neural network outputting output medical data in response to the complex-valued medical data being input to the neural network, and
wherein the linear calculation performed by the processing circuit includes a process in which a bias term is fixed to zero.
19 . A medical data processing method comprising:
applying, by a processing circuit, to medical data having a complex value, (1) a linear calculation that uses a complex number coefficient, and (2) a non-linear activation function of a neural network, wherein the activation function is not dependent on an argument of a complex number input into the neural network, the medical data being obtained by processing a signal received from a medical image diagnosis apparatus,
wherein the processing circuit includes the neural network having the non-linear activation function, the neural network outputting output medical data in response to the complex-valued medical data being input to the neural network, and
wherein the linear calculation performed by the processing circuit includes a process in which a bias term is fixed to zero.