Convolutional block attention gate-based Unet framework for microaneurysm segmentation using retinal fundus images.

backgroundDiabetic retinopathy is a major cause of vision loss worldwide. This emphasizes the need for early identification and treatment to reduce blindness in a significant proportion of individuals. Microaneurysms, extremely small, circular red spots that appear in retinal fundus images, are one of the very first indications of diabetic retinopathy. Due to their small size and weak nature, microaneurysms are tough to identify manually. However, because of the complex background and varied lighting factors, it is challenging to recognize microaneurysms in fundus images automatically.

methodsTo address the aforementioned issues, a unique approach for MA segmentation is proposed based on the CBAM-AG U-Net model, which incorporates Convolutional Block Attention Module (CBAM) and Attention Gate (AG) processes into the U-Net architecture to boost the extraction of features and segmentation accuracy. The proposed architecture takes advantage of the U-Net's encoder-decoder structure, which allows for perfect segmentation by gathering both high- and low-level information. The addition of CBAM introduces channel and spatial attention mechanisms, allowing the network to concentrate on the most useful elements while reducing the less relevant ones. Furthermore, the AGs enhance this process by selecting and displaying significant locations in the feature maps, which improves a model's capability to identify and segment the MAs.

resultsThe CBAM-AG-UNet model is trained on the IDRiD dataset. It achieved an Intersection over Union (IoU) of 0.758, a Dice Coefficient of 0.865, and an AUC-ROC of 0.996, outperforming existing approaches in segmentation accuracy. These findings illustrate the model's ability to effectively segment the MAs, which is critical for the timely detection and treatment of DR.

conclusionThe proposed deep learning-based technique for automatic segmentation of micro-aneurysms in fundus photographs produces promising results for improving DR diagnosis and treatment. Furthermore, our method has the potential to simplify the process of delivering immediate and precise diagnoses.