Transfer learning approach in deep neural networks for uterine fibroid detection

Convolutional neural network (CNN), a class of deep neural network, takes images as input and automatically extracts features for effective class prediction. The performance of CNN architecture is a major concern while dealing with fewer data. Traditional CNN architectures like ImageNet, AlexNet, GoogleNet are trained with a large quantity of data. Transfer learning is applied in order to combine pre-trained models such as Inception-V4 and support vector machine (SVM) for better performance with fewer data. The goal of the proposed approach is to efficiently detect fibroid presence in uterus MRI images. Features of these images are extracted using initial layers of Inception-V4 and are transferred to SVM during training. Several combinations on various network classifiers are tested and the performance metrics are evaluated. The proposed model attained an accuracy of 81.05% with a U-Kappa score of 0.402 on predicting fibroid images and 0.25% accuracy improvement compared with ResNet based model.