With the development of artificial intelligence and the large-scale application of computer aided diagnosis (CAD) system in clinical diagnosis, more and more disease screening technologies [1]–[2] can be realized without professional intervention. Because of the difficulties in obtaining images of labeled cervical lesions, there are few examples of application of vaginal endoscopic images in disease screening. In this paper, focusing on the imbalance of labelled image data, we consider the classification performance from the data level and algorithm level respectively. We use VGG19 network and transfer learning to classify cervical lesion images into cancer (ca), cervical intraepithelial neoplasia (cin), cervicitis (cv), HPV virus infection (hpv), normal (ok) and cervical polyp (polyp). At the data level, we use the SMOTE algorithm to oversample the minority classes. At the algorithm level, we use the weights balance of loss function to transform the network structure. These two methods are applied to VGG network respectively. Finally, the classification performance is evaluated by accuracy, confusion matrix, G-mean and F-score. Compared with directly feeding data into neural network, the SMOTE-TL method and DA-WB-TL method we proposed shows great superiority.
[1]
Luis Perez,et al.
The Effectiveness of Data Augmentation in Image Classification using Deep Learning
,
2017,
ArXiv.
[2]
Taghi M. Khoshgoftaar,et al.
A survey of transfer learning
,
2016,
Journal of Big Data.
[3]
Atsuto Maki,et al.
A systematic study of the class imbalance problem in convolutional neural networks
,
2017,
Neural Networks.
[4]
Nitesh V. Chawla,et al.
SMOTE: Synthetic Minority Over-sampling Technique
,
2002,
J. Artif. Intell. Res..
[5]
Xibei Yang,et al.
Recognition of Multiple Imbalanced Cancer Types Based on DNA Microarray Data Using Ensemble Classifiers
,
2013,
BioMed research international.
[6]
Sebastian Thrun,et al.
Dermatologist-level classification of skin cancer with deep neural networks
,
2017,
Nature.
[7]
Andrew Zisserman,et al.
Very Deep Convolutional Networks for Large-Scale Image Recognition
,
2014,
ICLR.