Cervical Cancer (CC) is the second most frequent malignancy in women worldwide, with a 60 % mortality rate; it is the leading cause of death worldwide. The majority of cervical cancer deaths occur in less developed countries where there is a lack of screening programs and sensitization about the disease. Because CC cannot be detected in its early stages since it does not reveal any symptoms and a long latent period. Accurate staging can aid radiologists in providing effective therapy by utilizing diagnostic methods such as MRIs. In this paper, two approaches are proposed, the first consist of introducing an automatic system for early detection of CC using image processing techniques and axial, sagittal T2-weighted MRIs for analysis to determine the pathological stage of tumour and to identify the real impact of cancer that will help the patient to be treated with high efficiency and properly. This detection process goes through three major steps, i.e. Preprocessing to make the representation of MRIs significant and easy to be analyzed, then the Segmentation was performed by Region Growing and Geometric deformable techniques to extract the Region Of Interests (ROIs).In the next step, we extract two categories of features based on Statistical and Transform methods in order to describe our ROIs, at the final step, five classifiers were trained to classify the MRIs into two classes: Benign or Malign. The second approach aims to increase the performance of pretrained Deep Convolutional Neural Networks (DCNNs) based on Transfer Learning (TL) used to classify our Female Pelvis Dataset (FP_Dataset) by adopting the stacking generalized method that provides a more efficient and robust classifier. Data augmentation is a pre-processing method applied to our MRIs and a dropout layer is used to prevent Networks from overfitting in our small Dataset. The results of experiments show that data augmentation and stacking generalization are an efficient way to improve accuracy rate of classification.
Key words: Cervical Cancer, MRI, Segmentation, Features, DCNNs, Transfer Learning, Stacking, Classification.
|
