Convolutional Neural Network with Multiscale Fusion and Attention Mechanism for Skin Diseases Assisted Diagnosis
Algorithm 1
LRAA.
Input:: Image data after preprocessing
Require:: The size of the convolution kernel, and
Require:: The distance between two pixels in transverse or longitudinal in a convolution kernel
Require:: Learning rate
Require:: is the previous loss value, and is the current loss value
Require:: Hyperparameters to control the decay rates for the momentum estimates
Require:: Hyperparameters to control the decay rates for learning rate
Require:: Loss function with set of all parameters, can be obtained by MSFA-NET, and
Require:: Initial parameter vector
: Initialize iteration
: Initialize moment vector
: Initialize moment vector
Output:: The output of the MSFA-NET
(1)
while not converged do
(2)
(3)
Encoder: Computing the result of five different ’s by (1)–(3); and results were obtained by (6) and (7) respectively; according to formulas (5)–(7), the result of is calculated by scSE (I) to obtain ; calculate the result after stitching through formula (4); and update through scSE (II)
(4)
ASCS: Calculate the result when by (5)–(7); by formula (8), the result of the Hadamard product of , and is obtained
(5)
Decoder: According to formulas (9)–(11), we get . By formula (12), the result of MSFA-NET is calculated
(6)
Loss: According to formulas (12)–(14), the loss function can be calculated
