Improving Brain Tumor Image Classification with Transfer Learning and Selective Augmentation

Brain tumors are increasingly prevalent health concerns, with approximately 90,000 people diagnosed with a primary brain tumor every year[i]. Glioma tumors are the most lethal and aggressive type of brain tumors and therefore there is a need for increased efficiency in diagnosis. Currently, radiologists diagnose brain tumors through Magnetic Resonance Imaging (MRI) scans. With the advent of machine learning technologies, the process of preliminary image classification can be optimized. The convolutional neural network (CNN), an example of machine learning implementations, is widely used for image classification. In this project, a baseline CNN model was developed using 3,264 labeled MRI images to classify between 4 categories: gliomas, meningiomas, pituitary tumors, and no tumors. The model’s initial overall accuracy was 74%. With image augmentation and transfer learning using EfficientNetB0, the accuracy improved to 80%. However, recall for the glioma category was only 33%, consistently lower than that of meningiomas, pituitary tumors and no tumors. By retraining the model with additional training images for only the glioma category, a method known as selective data augmentation, the recall for gliomas improved to 62%, and the model’s overall accuracy increased to 94%. To further investigate the low recall for glioma category, identifying potentially mislabeled training data and removing those images was also evaluated. Overall, the results indicate that transfer learning applied to CNN models can benefit diagnostic image classification. Selective augmentation and identifying noisy training data can be used to further improve performance.