Abstract:Feature fusion techniques have been proposed and tested for many medical applications to improve diagnostic and classification problems. Specifically, cervical cancer classification can be improved by using such techniques. Feature fusion combines information from different datasets into a single dataset. This dataset contains superior discriminant power that can improve classification accuracy. In this paper, we conduct comparisons among six selected feature fusion techniques to provide the best possible classification accuracy of cervical cancer. The considered techniques are canonical correlation analysis, discriminant correlation analysis, least absolute shrinkage and selection operator, independent component analysis, principal component analysis, and concatenation. We generate ten feature datasets that come from the transfer learning of the most popular pre-trained deep learning models: Alex net, Resnet 18, Resnet 50, Resnet 10, Mobilenet, Shufflenet, Xception, Nasnet, Darknet 19, and VGG Net 16. The main contribution of this paper is to combine these models and then apply them to the six feature fusion techniques to discriminate various classes of cervical cancer. The obtained results are then fed into a support vector machine model to classify four cervical cancer classes (i.e., Negative, HISL, LSIL, and SCC). It has been found that the considered six techniques demand relatively comparable computational complexity when they are run on the same machine. However, the canonical correlation analysis has provided the best performance in classification accuracy among the six considered techniques, at 99.7%. The second-best methods were the independent component analysis, least absolute shrinkage and the selection operator, which were found to have a 98.3% accuracy. On the other hand, the worst-performing technique was the principal component analysis technique, which offered 90% accuracy. Our developed approach of analysis can be applied to other medical diagnosis classification problems, which may demand the reduction of feature dimensions as well as a further enhancement of classification performance.

Fusion of Handcrafted and Deep Features for Medical Image Classification

Medical Image Fusion Using Transfer Learning and L‐BFGS Optimization Algorithm

Feature extraction of multimodal medical image fusion using novel deep learning and contrast enhancement method

Multimodal Medical Image Fusion: The Perspective of Deep Learning

Learning Based Combining Different Features for Medical Image Retrieval

Medical image fusion with deep neural networks

Multimodal medical image fusion and classification using deep learning techniques

Multi-Branch Cnn and Multi-Scale Multi-Dimensional Feature Fusion Mlp for Medical Image Classification

A Medical Image Fusion Method Based on SIFT and Deep Convolutional Neural Network in the SIST Domain

Deep learning of feature representation with multiple instance learning for medical image analysis

A review of deep learning-based information fusion techniques for multimodal medical image classification

Classification of Medical Images in the Biomedical Literature by Jointly Using Deep and Handcrafted Visual Features.

Hierarchical Progressive Network for Multimodal Medical Image Fusion in Healthcare Systems

Feature Extraction Processing Method of Medical Image Fusion Based on Neural Network Algorithm

MCFNet: Multi-Layer Concatenation Fusion Network for Medical Images Fusion

Deep Fusion of Shifted MLP and CNN for Medical Image Segmentation.

Deep Feature Engineering in Colposcopy Image Recognition: A Comparative Study

A review: Deep learning for medical image segmentation using multi-modality fusion

Adaptive spatial and frequency experts fusion network for medical image fusion

Medical Image Fusion: An insight into Deep Learning Approach

An Attention-based Multi-Scale Feature Learning Network for Multimodal Medical Image Fusion