The impact of region of interest size on brain tumor classification

Abstract

Differentiating between primary brain tumors and solitary brain metastases is crucial due to the different treatment approaches. The size of the region of interest (ROI) directly impacts the accuracy of computational analysis. This study evaluated the impact of 32 × 32 pixel ROIs on tumor classification, comparing the results to a previous study that used 10 × 10 pixel ROIs, aiming to contribute to the improvement of diagnostic techniques. Data were obtained from retrospective Magnetic Resonance Imaging exams. Sequences included Diffusion-Weighted Imaging (DWI), Fluid-Attenuated Inversion Recovery (FLAIR), T1-weighted with and without contrast, and T2-weighted. Forty texture parameters were extracted from each ROI using Matlab® and applied to machine learning methods: Naive Bayes, Support Vector Machine (SVM), Stochastic Gradient Descent (SGD), Random Forest, Decision Tree, and Neural Network. The 32 × 32 pixel ROIs showed the best performance on the T2 sequence, with an area under the curve (AUC) of 99.8% and an F1 score of 98.5%, both achieved with the Neural Network classifier. Compared to 10 × 10 pixel ROIs, the 32 × 32 pixel ROIs demonstrated superior performance and less variation among classifiers. Texture analysis of 32 × 32 pixel ROIs showed higher accuracy in distinguishing between primary brain tumors and solitary brain metastases, suggesting its potential as a non-invasive tool to improve clinical diagnosis in Magnetic Resonance Imaging.