Introduction

Theoretical background

Proposed method

Experimental results

Conclusion

References

Abstract

Super-resolution, which is one of the trend issues of recent times, increases the resolution of the images to higher levels. Increasing the resolution of a vital image in terms of the information it contains such as brain magnetic resonance image (MRI), makes the important information in the MRI image more visible and clearer. Thus, it is provided that the borders of the tumors in the related image are found more successfully. In this study, brain tumor detection based on fuzzy C-means with super-resolution and convolutional neural networks with extreme learning machine algorithms (SR-FCM-CNN) approach has been proposed. The aim of this study has been segmented the tumors in high performance by using Super Resolution Fuzzy-C-Means (SR-FCM) approach for tumor detection from brain MR images. Afterward, feature extraction and pretrained SqueezeNet architecture from convolutional neural network (CNN) architectures and classification process with extreme learning machine (ELM) were performed. In the experimental studies, it has been determined that brain tumors have been better segmented and removed using SR-FCM method. Using the SquezeeNet architecture, features were extracted from a smaller neural network model with fewer parameters. In the proposed method, 98.33% accuracy rate has been detected in the diagnosis of segmented brain tumors using SR-FCM. This rate is greater 10% than the rate of recognition of brain tumors segmented with fuzzy C-means (FCM) without SR.

Introduction

The brain has a complex structure consisting of millions of cells. Brain tumors are caused by the rapid growth of cells. Uncontrolled growth of cells affects brain activity and may damage normal cells [1]. Imaging technologies have played a role in the analysis of brain anatomy and functions with the development of technology. The tumor region can be detected by using Magnetic Resonance Imaging (MRI). However, radiologists need to measure the size of the tumor site for treatment. Image processing methodologies play an important role in the monitoring of tumor sites between radiologists and computers in the diagnosis and treatment process with machine learning. Radiologists improve diagnostic accuracy from a different perspective in interpreting medical images with these hybrid techniques [2]. Using a higher resolution on medical images makes it easier for the doctor to diagnose diseases. Super-resolution (SR) is used to increase the resolution. In the literature, SR is used in medical images [3–۵], MR images [6], brain MR images [6–۹], cardiac MR images [10], and retinal fundus images [11]. Many methods were used in the detection and automatic classification of tumor regions using MR images in the literature. Saad et al. [12] established a hybrid system with Fuzzy C-Mean (FCM) region growing for tumor analysis. Sachdeva et al. [13,34] classified the brain tumor by using the support vector machine (SVM) and artificial neural network (ANN) machine learning model in a hybrid structure with Genetic Algorithm (GA) (GA-SVM and GA-ANN). Mallick et al. [14] used a hybrid method with image compression of wavelet transform and image compression using Deep Wavelet Autoencoder (DWA) for brain MR images. In addition, images were classified using Deep Neural Networks (DNN). Performance comparison of DWA-DNN model with other conventional classification techniques was performed. Zeng et al. [9] suggest a deep convolutional neural network model for single and multiple contrast super-resolution reconstructions for artificial and real brain MR images. Swati et al. [15] suggest a method based on learning to transfer brain MR images using a pretrained deep CNN model. The proposed method has a coverage accuracy of 94.82% under five-fold cross-validation since it does not use any handmade features and requires minimal pretreatment.