AdaBoosted Deep Ensembles: Getting Maximum Performance Out of Small Training Datasets

摘要

Even though state-of-the-art convolutional neural networks (CNNs) have shown outstanding performance in a wide range of imaging applications, they typically require large amounts of high-quality training data to prevent over fitting. In the case of medical image segmentation, it is often difficult to gain access to large data sets, particularly those involving rare diseases, such as skull-based chordoma tumors. This challenge is exacerbated by the difficulty in performing manual delineations, which are time-consuming and can have inconsistent quality. In this work, we propose a deep ensemble method that learns multiple models, trained using a leave-one-out strategy, and then aggregates the outputs for test data through a boosting strategy. The proposed method was evaluated for chordoma tumor segmentation in head magnetic resonance images using three well-known CNN architectures; VNET, UNET, and Feature pyramid network (FPN). Significantly improved Dice scores (up to 27%) were obtained using the proposed ensemble method when compared to a single model trained with all available training subjects. The proposed ensemble method can be applied to any neural network based segmentation method to potentially improve generalizability when learning from a small sized dataset.