Attention-enriched deeper UNet (ADU-NET) for disease diagnosis in breast ultrasound and retina fundus images

In image segmentation, effective upsampling plays a pivotal role in recovering lost spatial information during the process of downsampling. Standard skip connections designed to mitigate this and prevalent in most models, often fall short of maintaining high segmentation performance, because of the loss of spatial information when it is transferred from the encoder to the decoder parts of the network. Responding to these limitations, we introduce the Attention-Enriched Deeper UNet (ADU-Net) a novel framework designed and tested on two medical image modalities – breast ultrasound images (BUSI) and retinal fundus images (RFI) for the effective transfer of information from the encoder to the decoder. The ADU-Net seamlessly combines global context modules (GCM) and progressive context refinement modules (PCRM) in the skip connection in a U-shaped network structure to capture the contextual information and enhance feature attention respectively. Another portion of the network, ConvGroup blocks set at its base facilitate the integration of convolutional operations while the deeper UNet was chosen for improved feature representation, handling of complex data, and enhanced hierarchical feature extraction. With these components, the ADU-Net effectively captures rich spatial details and crucial contextual information from input images resulting in precise and robust segmentation. ADU-Net achieved 76.49% and 59.19% mIoU (mean Intersection over Union) on BUSI and RFI datasets respectively, with 98.58% and 99.98% specificity. It also showed F1 scores of 62.11% and 71.71%, and accuracies of 94.87% and 96.02% on BUSI and RFI datasets respectively. ADU-Net is a powerful tool for medical professionals, aiding in the accurate localization of breast tumors, early detection of eye diseases, and formulation of optimal treatment plans. Its versatility allows it to analyze medical images across various modalities, making it suitable for diverse clinical scenarios.