The effect of spatial and intensity level augmentation of structural magnetic resonance images on autism diagnosis model

In deep learning, the robustness and generalizability of models significantly depend on diverse and heterogeneous training data. Acquiring such an extensive dataset is challenging in fields like disorder prediction due to data scarcity, which can be attributed to factors such as privacy concerns, limited patient population, or inadequate facilities. Data augmentation can be an ideal solution to this problem, particularly in the field of disorder prediction, like autism, using medical imaging. Data augmentation can expand and balance datasets by generating high-quality and varied data, thereby improving the generalizability of deep learning models. This study proposed two types of augmentation methods: 1. Spatial level 2. Intensity level augmentation techniques. Eight different levels of augmentations were experimented with across these categories. This study found that the combination of spatial and intensity level augmentations enhanced the model's generalizability and robustness, achieving an AUC value of 0.7433. Additionally, it was observed that the Left to Right flip method, under spatial augmentation, diminished the model's performance, whereas random noise injection, under intensity level augmentation, improved prediction accuracy.