HyperDense-Net: A densely connected CNN for multi-modal image segmentation
share
Neonatal brain segmentation in magnetic resonance (MR) is a challenging problem due to poor image quality and similar levels of intensity between white and gray matter in MR-T1 and T2 images. To tackle this problem, most existing approaches are based on multi-atlas label fusion strategies, which are time-consuming and sensitive to registration errors. As alternative to these methods, we propose a hyper densely connected 3D convolutional neural network that employs MR-T1 and T2 as input, processed independently in two separated paths. A main difference with respect to previous densely connected networks is the use of direct connections between layers from the same and different paths. Adopting such dense connectivity leads to a benefit from a learning perspective thanks to: i) including deep supervision and ii) improving gradient flow. This approach has been evaluated in the MICCAI grand Challenge iSEG and obtains very competitive results among 21 teams, ranking first and second in many metrics, which translates into a promising performance.
READ FULL TEXT
