The effect of input size on the accuracy of a convolutional neural network performing brain tumor detection

Abstract

A brain tumor can negatively affect basic bodily functions and when malignant can result in low survival rates. Many studies were conducted to detect and classify brain tumors in MRI images using a convolutional neural network (CNN) and other techniques like image preprocessing and transfer learning. However, few studies have explored the effect of specific hyperparameters on the performance of such CNNs. This study aims to investigate how the input size affects the CNN’s accuracy in brain tumor detection. Brain MRI datasets were collected and split into training, validation, and test sets. Four models with identical architectures but different input sizes of 256px×256px×3, 224px×224px×3, 128px×128px×3, and 64px×64px×3 were built using TensorFlow Keras, trained on the training set with data augmentation, and evaluated using the test sets. Of these four models, the one with 64px as input size has the best performance, yielding the highest test accuracy, 99.16%, and lowest test loss, 0.0282, whereas the 224px model has the worst performance, with the lowest accuracy, 98.06%, and highest loss, 0.0976. Accordingly, it appears that larger input sizes do not necessarily result in higher accuracy of the CNN performing brain tumor detection. Future studies on this topic may consider using a smaller input size, not only maintaining high accuracy but also significantly reducing the required time to train and the space to save the model.