Differentiation of Low and High-Grade Meningiomas Using Transfer Learning on MR Images

Despite the advantages of magnetic resonance (MR) images in diagnosing intracranial tumours such as meningiomas, non-invasive grading remains challenging. In this respect, the potential of transfer learning and data augmentation methods was explored to develop computer-aided diagnosis (CAD) systems for automatically differentiating between low and high-grade meningiomas on MR images. Four MR modalities were enrolled, including T1-weighted, T2-weighted, fluid-attenuated inversion recovery and T1-weighted contrast enhanced and the effectiveness of different data augmentation and deep learning approaches in determining the optimum AI-based solution for meningioma grading was investigated using a multi-stage framework, which included image preprocessing, data augmentation by either conventional geometric transformations or fancy principal component analysis (PCA) and the use of ImageNet pre-trained, well-known convolutional neural networks (CNNs) for training models with the two widely utilised optimisers of adaptive moment estimation (Adam) and stochastic gradient descent (SGD). Finally, performance was evaluated utilising standard metrics to select the best-obtained model with optimal approaches. Results indicated that trained visual geometry group (VGG)-19 with generated images by fancy PCA performed best with an accuracy of 97.80% and 98.90% in classifying unseen samples using SGD and Adam optimisers, respectively. This procedure also yielded acceptably reproducible results in a second study, ensuring the efficiency of each model using a small-scale dataset. Developing CAD systems using pre-trained CNNs and fancy PCA is a promising approach for classifying meningiomas on MR images into low and high-grade categories, which can serve as an alternative to invasive methods and provide valuable assistance for accurate diagnosis before treatment.