Machine Learning-Based Classification of Brain Tumors Using Perceptron Model and Nuclear Magnetic Relaxometry Data

Abstract

A brain tumor is an abnormal growth of cell that develops in the brain; this can be malignant or noncancerous. The classifications of these tumors are glioblastomas, low grade gliomas and metastases. The region of interest are contralateral white matter CW, peritumoral white matter PW, and solid tumor ST. Conventional magnetic resonance imaging (MRI) scanners are one of the image modalities used for brain tumor diagnoses. However, conventional MRI scans are not automated thereby making diagnosis difficult and provides inaccurate result, time-consuming and expensive. The literature has not yet determined how relaxometry investigations with machine learning models and computational techniques can distinguish between intra-axial brain tumors. The perceptron neural networks (PNN) provides a powerful tool to help doctors analyze, model and make sense of complex clinical data with quantitative parameters such as longitudinal and transverse relaxation times (T1 and T2) respectively. The study provides and creates machine learning models in relaxometry studies in which diagnosis can be done without undergoing full MRI scans and these processes are based on short time acquisition. The study also helps to check patient's treatment performance especially those who are not allergic to MRI contrast agent and help in follow-up studies. The Bloch equation was used to simulate the acquisition parameters (T1 and T2) in R programming language. A code in R-studio was used to extract the mean ± standard deviation (SD) data from 31 patients with untreated intra-axial brain tumors in another study, which were randomly generated from 31 to 900 hypothetical patients in, excel to have large dataset as sample for training models. In machine learning classification, negative binomial (nb) had the best performance model for CW with the accuracy of 79% and 68% for kappa, while support vector machine (SVM) model had the best model for PW with accuracy of 70% and 55% for kappa and glmnet model had the best performance model for ST with accuracy of 72% and kappa is 58%. These models with higher accuracy and kappa were used for deployment in machine implementation on CW, PW and ST. With these quantitative magnetic resonance imaging (MRF) method, with relaxometers, software, and models that allow for the analysis of results in less than five minutes, this study looks into the issue of time-consuming and sufficiently reduced MRI scan costs as well as automation for accurate diagnosis, making it faster, less expensive, and more efficient.