Trifuse: Triplet Encoders Network for Medical Image Segmentation

Deep learning techniques have recently played a crucial role in medical image segmentation. However, convolution-based deep learning models are limited by their local receptive fields, making it challenging to capture long-range dependencies. While recent models, such as Transformer and Mamba effectively process long sequences, they treat images as rigid structures. To address this limitation, we propose an efficient neural network architecture, TriFuse, which combines three distinct encoders. The EfficientNet encoder captures local information, the Mamba (S6) encoder captures global context, and a third Graph Neural Network (GNN) encoder, GreedyViG, captures irregular and complex structures. These three encoders are fused through a modified BiFusion module, and deep supervision using combinatorial aggregation is applied for loss calculation. Our evaluation on two datasets demonstrates superior performance, as measured by DSC and HD95.