Authors
-
Oluwafemi Williams Eweje
Department of Computer Science, University of Ibadan, Nigeria
-
Oluwashola David Adeniji
Department of Computer Science, University of Ibadan, Nigeria
-
Solomon Olalekan Akinola
Keywords:
Pooling, Black hat morphology, Otsu binarisation, Balanced multiclass accuracy
Abstract
Deep Convolutional Neural Networks (DCNN) involve alternating convolutional layers, non-linearity layers and pooling layers for identifying patterns in input. The pooling retains important information while down sampling the dimensionality of the feature map on dermoscopic images used for early cancer diagnosis. Existing DCNN for dermoscopic image analysis employs Max Pooling (MP) and Average Pooling (AP) due to their efficiency. The MP works best on images of dark background with lighter object, while AP works better on images of lighter background with darker object. An online International Skin Imaging Collaboration (ISIC) dermoscopic image dataset obtained from 2016 - 2019 was used for the research. A novel DCNN, IP-DCNN developed and configured with rectified linear unit activation function, multiclass cross entropy loss and Softmax functions.
Evaluation of the IP-DCNN with filtered-segmented images was done by comparing its performance with existing current studies which used DCNN architectures. The developed interpool deep convolutional neural networks provided an improved performance over the pure deep convolutional neural networks and its existing variants.
Author Biography
Oluwafemi Williams Eweje, Department of Computer Science, University of Ibadan, Nigeria
Deep Convolutional Neural Networks (DCNN) involve alternating convolutional layers, non-linearity layers and pooling layers for identifying patterns in input. The pooling retains important information while down sampling the dimensionality of the feature map on dermoscopic images used for early cancer diagnosis. Existing DCNN for dermoscopic image analysis employs Max Pooling (MP) and Average Pooling (AP) due to their efficiency. The MP works best on images of dark background with lighter object, while AP works better on images of lighter background with darker object. An online International Skin Imaging Collaboration (ISIC) dermoscopic image dataset obtained from 2016 - 2019 was used for the research. A novel DCNN, IP-DCNN developed and configured with rectified linear unit activation function, multiclass cross entropy loss and Softmax functions. Evaluation of the IP-DCNN with filtered-segmented images was done by comparing its performance with existing current studies which used DCNN architectures. The developed interpool deep convolutional neural networks provided an improved performance over the pure deep convolutional neural networks and its existing variants.
