HyperDense-Net: A Hyper-Densely Connected CNN for Multi-Modal Image Segmentation.

Area of Science:

• Computer Vision
• Medical Image Analysis
• Deep Learning

Background:

• Dense connections in neural networks improve gradient flow and performance in computer vision.
• Existing multi-modal Convolutional Neural Network (CNN) approaches often fuse modalities at a single network level, limiting representation learning.

Purpose of the Study:

• To introduce HyperDenseNet, a 3-D fully convolutional neural network extending dense connectivity to multi-modal segmentation.
• To enable learning complex combinations between imaging modalities at all levels of abstraction.

Main Methods:

• Proposed a 3-D fully convolutional neural network architecture with dense connections within and across imaging modality paths.
• Evaluated the network on the iSEG 2017 (infant brain) and MRBrainS 2013 (adult brain) multi-modal segmentation challenges.

Main Results:

• HyperDenseNet achieved significant improvements over state-of-the-art segmentation networks on both benchmark datasets.
• The network ranked at the top in both the iSEG 2017 and MRBrainS 2013 challenges.
• Experimental analysis confirmed the importance of hyper-dense connections for multi-modal representation learning.

Conclusions:

• HyperDenseNet effectively leverages dense connections for enhanced multi-modal brain tissue segmentation.
• The proposed architecture offers a more flexible and powerful approach to multi-modal fusion in deep learning for medical imaging.