Multi-scale Ensemble of ResNet Variants

Residual learning has become a staple in the deep learning community due to its simple yet effective design. ResNets have been successfully employed for a variety of problems (Tan et al. 2018; Chen et al. 2018; Habibzadeh et al. 2018; Putten et al. 2019). Additionally, we incorporate multi-scale information in our approach by training models with different input image resolutions. This approach is taken since multi-scale approaches have been shown to be effective for many medical image analysis problems (Litjens et al. 2017). Finally, ensembling is a good way to boost performance and ensembles have been used to win many AI competitions. These methods are especially effective when the models are diverse (Brown et al. 2005). We achieve this diversity by using different ResNet models and by employing the multi-scale approach.