MACFNet: Multi-attention cross-scale fusion network for OD and OC segmentation

Transformers show promise for segmenting the Optic Disc (OD) and Optic Cup (OC) from fundus images, offering potential advancements in automated glaucoma detection. However, traditional patch-based or token-based image processing often compromises essential channel-specific and spatial information. This results in an insufficient representation of key features and spatial relationships, ultimately diminishing segmentation accuracy. To address these challenges, a Multi-attention Cross-scale Fusion Network (MACFNet) is proposed. Specifically, MACFNet takes a Self-attention Convolution (SC) module as the encoder-decoder, synergizing the strengths of convolution and Transformer architectures while mitigating the computational burden associated with self-attention operations. Moreover, the Dual Attention (DA) modules are also incorporated into the encoder, which focuses on critical regions and features of medical images, thereby fully leveraging channel and positional feature information. Furthermore, a Multi-scale Cross Attention (MSCA) module is designed for the skip connection process, which effectively bridges the semantic gap of different stages by integrating global semantic features with local features. Experimental evaluations on two public datasets, DRISHTI-GS1 and REFUGE, achieving state-of-the-art performance in OD and OC segmentation tasks with the highest IoU and Dice scores compared to existing methods.