Issuu

International Research Journal of Engineering and Technology (IRJET)

e-ISSN: 2395-0056

Volume: 11 Issue: 08 | Aug 2024

p-ISSN: 2395-0072

www.irjet.net

COMPARATIVE ANALYSIS OF DEEP LEARNING TECHNIQUES FOR BRAIN TUMOR DETECTION Spoorthy UK1, Madhuri J2 1Student BIT Dept of CSE Bangalore, India

2Assistant Professor BIT Dept of CSE Bangalore, India

-----------------------------------------------------------------------***------------------------------------------------------------------ Validate Model Generalizability Ensure robust Abstract—The project classifies brain tumors from MRI performance across diverse data sets with rigorous validation techniques.

images using advanced deep learning. It employs CNNs with attention mechanisms and GANs for data augmentation. The dataset includes glioma, meningioma, pituitary tumors, and no tumors. DenseNet and custom Attention CNN improve detection accuracy. This assists medical professionals in early diagnosis and treatment planning.

 Facilitate Clinical Integration Design the system for easy

adoption in existing medical workflows.  Contribute

to Medical Research Provide actionable insights to advance medical imaging and oncology.

Keywords—Convolutional Neural Networks (CNNs),

II. LITERATURE SURVEY

Generative Adversarial Networks (GANs), DenseNet, Attention CNN

Anurag Goswami, Manish D.[3], aims to replace the missing parts of the facial photos using Recurrent Generative Adversarial Network. Feature Representation Net also regarded as the semantic feature extractor. Three feature extractors with multiple scales are in use. The use of dilated convolution has increased the receptive field. Feature Transformation Net regarded as mapping function over the feature domain. The input is the feature extracted from FTN and output will be transferred features. This is able to be rebuilt face images. Here ConvLSTM has been adopted. The image reconstruction net predicts the facial picture at each scale by using the characteristics produced by the FTN. Despite the fact that there are many of options for upsampling or downsampling, they adopt the simple bilinear interpolation for all the experiments in their work. To fully exploit the correction of adjacent level features of their model, they design a novel short link structure to fuse the multi-scale features. Discriminator network has global and local discriminator. Zhang, Mingming et al.[5], aims to construct an improved Generative Adversarial Networks based on the context encoder and proposes a selflocalization occlusion method for restoring faces in images algorithm. The generator in this paper adopts the convolution network with the structure of Variational Autoencoder. The encoder uses 12 convolutional layers , 1 fully connected layer. The decoder uses 12 deconvolutional layer and one fully connected layer. Leaky ReLu is used as activation function in the first 25 layers and Tanh is used as activation in the last layer. The discriminator in this paper is based on VGG19 with 13 convolutional layer and 5 pooling layers. CelebA was the dataset used, and the training Adam optimizer had a learning rate of 0.002. Jiang, Yi, Jiajie Xu et al.[8], presents an approach for face image inpainting using skip connection layers between the encoder and the decoder. Inpainting is part of a large set of image generation problems. To solve this problem, an

I. INTRODUCTION Detecting and diagnosing brain tumors through MRI scans is critical for patient care and treatment planning. Traditional methods rely heavily on radiologists, which can be time-consuming and prone to human error. Deep learning offers a promising alternative, with CNNs, GANs, DenseNet, and Attention Mechanisms enhancing detection accuracy. This project uses a comprehensive MRI dataset categorized into glioma, meningioma, pituitary tumors, and no tumors. Models, including DenseNet for its feature extraction capabilities and Attention CNN to highlight relevant image parts, are trained on this data. GANs are used to generate synthetic images, addressing data scarcity and class imbalance. The goal is to develop a highly accurate and reliable model that assists medical professionals in diagnosing brain tumors, improving patient outcomes, and streamlining healthcare services. By integrating these advanced techniques, the project aims to enhance the accuracy, efficiency, and interpretability of brain tumor detection. A. Objectves:  Enhance

Detection Accuracy Develop a model to significantly improve tumor detection accuracy.

 Increase Diagnostic Efficiency Streamline the diagnostic

process to reduce time and resource consumption.  Improve

Model Interpretability Implement tools like Grad-CAM to make model decisions transparent.

 Address

Data Limitations Use GANs for dataset augmentation to overcome data scarcity.

Impact Factor value: 8.226

ISO 9001:2008 Certified Journal

Page 379