A Novel ConvLSTM-Based U-net for Improved Brain Tumor Segmentation

Using 2D scans or simple 3D convolutions are two limitations of previous works on segmentation of brain tumors by deep learning, which lead to ignoring the temporal distribution of the scans. This study proposes a novel extension to the well-known U-net model for brain tumor segmentation, utilizing 3D Magnetic Resonance Imaging (MRI) volumes as inputs. The method, called ConvLSTM-based U-net + up skip connections, incorporates the ConvLSTM blocks to capture spatio-temporal dependencies in the 3D MRI volumes, and up skip connections to capture low-level feature maps extracted from the encoding path, enhancing the information flow through the network to the standard U-net architecture. A novel intensity normalization technique is used to improve the comparability of scans. This technique normalizes image intensity by subtracting the grey-value of the most frequent bin from the image. The novel method is tested on the Multimodal Brain Tumor Segmentation (BRATS) 2015 dataset, showing that the use of ConvLSTM blocks improved segmentation quality by 1.6% on the test subset. The addition of skip connections further improved performance by 3.3% and 1.7% relative to the U-net and ConvLSTM-based U-net models, respectively. Moreover, the inclusion of up skip connections could enhance the performance by 5.7%, 3.99% and 2.2% relative to the simple U-net, ConvLSTM-based U-net, and ConvLSTM-based U-net with skip connections, respectively. Finally, the novel preprocessing technique had a positive effect on the proposed network, resulting in a 3.3% increase in the segmentation outcomes.