Brain Tumor Segmentation Using Two-Stage Convolutional Neural Network for Federated Evaluation

A deep learning method is proposed for brain tumor segmentation using a two-stage encoder-decoder convolutional neural network (CNN). To improve the generalization of the proposed network for federated evaluation, we propose a two-stage encoder-decoder CNN that performs coarse segmentation at stage-I and fine segmentation at stage-II. Stage-I consists of an ensemble of three predictions on the orthogonal slices of a subject. In stage-II, the predictions of the first stage are used to crop the region of interest consisting of the tumor region and a fine grain segmentation is performed on the cropped image. A single ResUNet was used for stage-I and seven different networks were used for stage-II. Heavy data augmentation consisting of geometric transformation and random contrast was used to avoid overfitting and improve the generalization. The mean dice scores on 21 imaging sites evaluated in a federated manner achieved dice scores of 0.8659, 0.7708, and 0.7714 for the whole tumor, tumor core, and enhancing tumor respectively. The method ranked second in the federated evaluation task.