Enhanced Mitochondrial Segmentation with Customized Attention U-Net and Nakagami Fusion

Accurate and efficient segmentation of mitochondria from electron microscopy images is crucial for a wide range of biological and medical applications. In this paper, we present a novel approach to improve mitochondrial segmentation by introducing an added Nakagami parametric (ANP) image-based preprocessing technique. We combine this preprocessed image with the original input that enhances the robustness and generalization of the segmentation model. To find the optimal combination of methods, we explore three different loss functions to train our model for varying epochs. We adopt the Attention U-Net architecture, customized it according to our strategy and demonstrated its superior performance in terms of mean Intersection over Union (IoU) and mean Dice score when compared to state-of-the-art methods. Our results showcase the potential of our approach for improved mitochondrial segmentation, offering significant contributions to biomedical image analysis and related fields.