Nakagami Weighted Parametric Image-Based Segmentation of Microscopy Images Using U-Net

This study presents an innovative approach for segmenting medical microscopy images, integrating the Nakagami parametric distribution with the U-Net convolutional neural network. Nakagami distribution models the brain tissue's hippocampus region scattering properties, aiding mitochondria segmentation. U-Net, an ideal choice for medical image segmentation, is trained to discern complex patterns within brain tissue and mitochondria regions. Evaluation on a publicly available electron microscope image dataset highlights its superiority over current methods, excelling in pixel accuracy, precision, recall, mean intersection over union (mIoU), and dice coefficient score. The approach is computationally efficient, optimized for low-configuration GPUs, enabling realtime clinical applications. The customized U-Net architecture promises lightweight and faster computation, potentially enhancing diagnosis and treatment outcomes. Experimental results reveal impressive performance, with dice coefficient score (DCS), mIoU, pixel accuracy (pACC), precision (PPV), and recall (sensitivity) at 93.83%,93.01 %,97.82%,95.29%, and 93.04%, respectively, surpassing conventional machine learning and deep learning methods.