TY - JOUR
T1 - A Multilayer Perceptron Based Smart Pathological Brain Detection System by Fractional Fourier Entropy
AU - Zhang, Yudong
AU - Sun, Yi
AU - Phillips, Preetha
AU - Liu, Ge
AU - Zhou, Xingxing
AU - Wang, Shuihua
N1 - Publisher Copyright:
© 2016, Springer Science+Business Media New York.
PY - 2016/7/1
Y1 - 2016/7/1
N2 - This work aims at developing a novel pathological brain detection system (PBDS) to assist neuroradiologists to interpret magnetic resonance (MR) brain images. We simplify this problem as recognizing pathological brains from healthy brains. First, 12 fractional Fourier entropy (FRFE) features were extracted from each brain image. Next, we submit those features to a multi-layer perceptron (MLP) classifier. Two improvements were proposed for MLP. One improvement is the pruning technique that determines the optimal hidden neuron number. We compared three pruning techniques: dynamic pruning (DP), Bayesian detection boundaries (BDB), and Kappa coefficient (KC). The other improvement is to use the adaptive real-coded biogeography-based optimization (ARCBBO) to train the biases and weights of MLP. The experiments showed that the proposed FRFE + KC-MLP + ARCBBO achieved an average accuracy of 99.53 % based on 10 repetitions of K-fold cross validation, which was better than 11 recent PBDS methods.
AB - This work aims at developing a novel pathological brain detection system (PBDS) to assist neuroradiologists to interpret magnetic resonance (MR) brain images. We simplify this problem as recognizing pathological brains from healthy brains. First, 12 fractional Fourier entropy (FRFE) features were extracted from each brain image. Next, we submit those features to a multi-layer perceptron (MLP) classifier. Two improvements were proposed for MLP. One improvement is the pruning technique that determines the optimal hidden neuron number. We compared three pruning techniques: dynamic pruning (DP), Bayesian detection boundaries (BDB), and Kappa coefficient (KC). The other improvement is to use the adaptive real-coded biogeography-based optimization (ARCBBO) to train the biases and weights of MLP. The experiments showed that the proposed FRFE + KC-MLP + ARCBBO achieved an average accuracy of 99.53 % based on 10 repetitions of K-fold cross validation, which was better than 11 recent PBDS methods.
KW - Biogeography-based optimization
KW - Fractional Fourier entropy
KW - Multilayer perceptron
KW - Pathological brain detection system
KW - Pruning
KW - Real-coded
UR - http://www.scopus.com/inward/record.url?scp=84976314220&partnerID=8YFLogxK
U2 - 10.1007/s10916-016-0525-2
DO - 10.1007/s10916-016-0525-2
M3 - Article
C2 - 27250502
AN - SCOPUS:84976314220
SN - 0148-5598
VL - 40
JO - Journal of Medical Systems
JF - Journal of Medical Systems
IS - 7
M1 - 173
ER -