•  
  •  
 

Abstract

Artificial intelligence, especially in the field of ``deep learning'', is still promising when it comes to skin cancer detection and diagnosis. Among deep learning algorithms, convolutional neural networks (CNNs) give a high level of accuracy in identifying and classifying different types of skin cancer. CNNs have a strong coordination due to understanding the important features from medical images that are extracted from convolutional layers. However, there is still a problem which is the high imbalance in the dataset with high noise in the images. This paper presents a new solution that combines different architectural structures of convolutional neural networks (CNNs) to extract features from skin cancer images with two forms of linear algebra methodologies, Principal Component Analysis (PCA) and Factor Analysis (FA) to eliminate the curse of combining the extracted features and removing the redundant features and passing them to a fully connected classifier to perform multiple classification, in this paper the HAM10000 dataset is used. The idea here is to retain the vital features of the image and extract the important features from it and eliminate the irrelevant features which leads to achieving higher classification accuracy for all classes. The proposed method was accurate and achieved a macro average test accuracy of 97.90% with a test loss of 0.14% and precision, recall and f1 score of 0.97%, 0.93% and 0.95% respectively and achieved a weighted average of 0.98% in precision, recall and f1 score in multi-class classification.

Download

Share

COinS