Abstract
A brain tumor is a dangerous cancer that develops when cells divide uncontrollably and abnormally. Recent advancements in deep learning have aided the medical imaging industry in diagnosing various disorders medically. Convolutional neural networks are the most often used machine learning algorithm for visual recognition and learning. Additionally, we demonstrate by using CNN to classify brain MRI images into two categories: cancer and non-cancer. Using the transfer learning method, we evaluated our convolutional model’s performance to previously trained ResNet-v2-152, Inception-v3, and Inception-Resnet-v2 models. As a result of the experiment, a moderate dataset was used. However, the test result indicates that the suggested model’s accuracy was adequate, reaching 99 percent, compared to 98 percent for ResNet-v2-152, 98 percent for Inception-v3, and 97 percent for Inception-Resnet-v2. The suggested model requires far less computational resources and is more efficient.
Introduction
Brain tumors are caused by an unusual growth of cells in the brain and can either be primary (non-cancerous) or secondary (cancerous). Primary tumors are benign and don’t spread, but secondary tumors can spread to other areas of the brain and body. As a tumor grows, it increases pressure on the brain, which can be dangerous. Early detection of brain tumors is important for successful treatment. Brain tumors are considered to be one of the most serious and deadly forms of cancer that both adults and children can get. However, if treated early and properly, it can prevent the disease from spread, stop the growth of tumors, and reduce the cost of medicine and treatment. Today, specialists can diagnose brain tumors using computer systems with modern technology, making it easier for professionals to identify tumors while avoiding errors in traditional methods. In the past, biopsy was used for diagnosis but it was uncomfortable for patients and required prior MRI or CT scans. MRI is a safer option as it doesn’t use radiation and provides precise imaging of malignancies. However, manual examination of many MRI images is time-consuming and not always accurate. Automated detection using convolutional neural networks has shown to provide more accurate results than traditional methods, as radiologists can miss between 10-30% of cancers during both diagnostic and screening stages.
Related work
Abdul Hannan Khan and et al., [1] presents a hierarchical deep learning (HDL2BT) model, leveraging convolutional neural networks (CNNs), for tumor classification into four categories: glioma, meningioma, pituitary, and no tumor. The model demonstrated an accuracy of 92.13% and an error rate of 7.87%. Sidra Sajid and et al., [2] propose a method that involves a pre-processing stage for normalizing the images and correcting bias field, followed by a feed-forward pass through a CNN, and a post-processing stage to eliminate false positives around the skull region. The proposed method is evaluated on the BRATS 2013 dataset, attaining DICE score, sensitivity, and specificity scores of 0.86, 0.86, and 0.91 for the whole tumor region, respectively.
Masoumeh Siar and Mohammad Teshnehlab., [3] proposed a model that utilizes a Radial Basis Function (RBF) classifier and a Decision Tree (DT) to obtain an accuracy of 97.34% and 94.24% respectively in a Convolutional Neural Network (CNN). The results indicate that the Softmax classifier achieved the highest accuracy within the CNN. The proposed method demonstrated an accuracy of 99.12% on the test data. Tanzila Saba and et al., [4] proposed a method that uses the Grab Cut technique to precisely segment real lesion marks, utilizing a transfer learning model (VGG-19). The method was tested using the Dice similarity coefficient (DSC) on the BRATS 2015, 2016, and 2017 datasets, resulting in scores of 0.99, 1.00, and 0.99, respectively.
Hari Mohan Rai and Kalyan Chatterjee,. [5] proposed a method for abnormality detection from brain MR images using the U-Net model architecture with less and less complex layers (LeU-Net) is proposed. The LeU-Net model results overall record of 98% accuracy on cropped images and 94% accuracy on uncropped images. Zheshu Jia and et al., [6] presented a Fully Automatic Heterogeneous Segmentation using Support Vector Machine (FAHS-SVM) method for brain tumor segmentation utilizing deep learning techniques. The numerical results demonstrate an accuracy of approximately 98.51% in separating abnormal and normal tissue in brain magnetic resonance images.
Ali Mohammad Alqudah and et al., [7] applied a Convolutional Neural Network (CNN) to classify brain tumors into three classes (glioma, meningioma, and pituitary tumor) using a dataset of 3064 T1 contrast-enhanced brain MR images. The model showed an overall accuracy of 98.93% and sensitivity of 98.18% for cut lesions and 99% accuracy and 98.52% sensitivity for uncut lesions. For segmented lesion images, the model achieved an accuracy of 97.62% and sensitivity of 97.40%. Nadim Mahmud Dipu and et al., [8] used the Yolo and FastAi libraries to apply it to the BRATS 2018 dataset. The results showed that the YOLOv5 model achieved an accuracy of 85.95%, while the FastAi classification model achieved an accuracy of 95.78%.
Neelum Noreen and et al., [9] investigated the use of two pre-trained deep learning models, Inception-v3 and DensNet201, for tumor detection. The results showed that the models achieved a test accuracy of 99.34% and 99.51%, respectively. P Gokila Brindha and et al., [10] proposed the use of a self-defined artificial neural network (ANN) and a convolutional neural network (CNN) for brain tumor detection and analyzed their performance. The ANN model generated in this work achieved a test accuracy of 65.21%. Ahmad Saleh and et al., [11] trained a brain tumor dataset using five pre-trained models: Xception, ResNet50, InceptionV3, VGG16, and MobileNet. The F1 score measurement of unseen images achieved 98.75%, 98.50%, 98.00%, 97.50%, and 97.25% respectively.
Suggested techniques
The research evaluated 3,000 brain MRI scans (Figure. 3-1) by using image processing techniques. A six-layer convolution model was employed for training purposes and its performance was compared to pre-existing models such as ResNet-v2-152, Inception-v3, and Inception-Resnetv2. The dataset used consisted of 1,500 images of malignant cancers and another 1,500 images of benign non-cancerous tumors.