Artificial Intelligence-Based Classification of Brain Tumors Using Garson Model and Magnetic Resonance Fingerprinting

Abstract

Conventional magnetic resonance imaging (MRI) scans are not automated thereby making diagnosis difficult, time-consuming and expensive. In low- and middle-income countries (LMICs), these scanners are inadequate and the few available ones are not regularly maintained due to poor funding. Consequently, patients with intra-axial brain tumors are mostly not diagnosed early enough for treatment. Furthermore, enough scanners are not available for follow up assessment after treatment and worse still; most patients avoid follow up assessment due to the cost of having an MRI scan. To address this challenge, this study assessed the utility of magnetic resonance fingerprinting (MRF) and machine learning (ML) for differentiating common types of adult intra-axial brain tumors (gliomablastomas, low grade gliomas and metastases). In this study, emphasis was placed on the use of two parameters measured with MRF, T1 and T2 relaxations times. T1 and T2 values for these brain tumors were obtained from literature for 31 patients with untreated intra-axial brain tumors and scanned using magnetic field strength (B0) of 1.5T. The three classes of tumors were considered in three regions of interests (ROI); that is, contralateral white matter (CW), peritumoral white matter (PW) and solid tumor (ST). A code in R-studio was used to extract the mean ± standard deviation (SD) of T1, T2 data from 31 patients with untreated intra-axial brain tumors in another study, since the data from 31 patients is quite small to train ML models, a larger data was simulated with a computer program written in R in which 300 patients were assumed to be scanned. In all, 900 dataset (representing each patient) were simulated such that each tumor class had 300 data points each. The time-independent Bloch nuclear magnetic resonance (NMR) flow equation was then solved analytical for transverse magnetisation (M). Using the simulated data points, the relaxation rate (T0) and M were then calculated for each tumor class. The calculations were then pre-processed into a dataset for training ML models. An artificial neural network such as Garson Algorithm was implemented for the dataset. The results showed the model was able to identify complexities in the inputs such that T0 was found to be the most important feature determining MR signal for CW and PW while T1 is most important feature determining the signal. 80% of this dataset was used as training set while the remaining 20% was used as testing dataset for implementation of supervised ML models. nb had the best classification accuracy of 79.17% in CW, svmR produced the best classification accuracy of 70.42% in PW while glmnet had the best accuracy of 72.50% in ST. The best performing models were then deployed as an interactive application to demonstrate how this study can be used for clinical diagnosis of MRF-based diagnosis of brain tumors. The Garson model was used to determine the relative important of input variables in a neural network by examine the connection weight approach. The model could also be used to provide the results of a prediction model, which displays the connection between the predictor and outcomes after modified parameters in the same model; however, it may not provide a direct dependent interpretation.