Content Details | NUML Online Research Repository

Title MRI BASED BRAIN TUMOR CLASSIFICATION AND DETECTION THROUGH MULTI-MODAL DEEP LEARNING TECHNIQUES

Author(s) Saima Razzaq

Abstract Machine learning and deep leaning methods have substantially advanced the efficacy of disease diagnosis in healthcare setups, by facilitating precise and early prognosis of disease, enabling timely intervention and resource optimization. One of the key field where it is stated beneficial is brain tumor diagnosis, a most serious disease which can adversely impact the people of any age group. Despite of significant advancements in the field of deep learning in visual data processing there are still challenges. The research highlighted important challenges in brain tumor scrutiny; comprising morphological uncertainty, tumor heterogeneity, class imbalance, data scarcity, and model accuracy. Besides these challenges the processing of various medical imaging modalities; i.e the data which is obtained from different medical devices like MRI, CT, and PET have inconsistent features, the accuracy results are variant and inadequate. By considering these factors, the research is based on development of two deep learning models; the LSTM model and hybrid 2D UNET + LSTM model to accurately identify and classify four types of brain tumor. Firstly, MRI images are preprocessed using N4ITK bias field correction to eliminate intensity inhomogeneties and to boost their qualities. Then the proposed Hybrid model combine the 2D UNET and LSTM networks, with 2D UNET having four convolutional blocks with variable number of 3x3 sized filters. The networks and layers arrangements are chosen after extensive experimentation. Model performance was accessed using various key metrics including precision, recall, F1 score and specificity. The results demonstrate a significant accuracy of 99.12% of hybrid 2D U-Net + LSTM which outperforms the standalone LSTM model. Additionally, comparative analysis with existing research is performed which notices the better outcomes of the proposed models. Therefore, the research helps in advancing tumor classification accuracy and reliability in medical research

Type Thesis/Dissertation MS

Faculty Engineering and Computer Science

Department Engineering

Language English

Publication Date

Subject

Publisher

Contributor(s)

Format

Identifier

Source

Relation

Coverage

Rights