Abstract
This study investigates the integration of attention mechanisms within deep learning architectures, specifically focusing on the segmentation of diplococci microorganisms in microscopy images. We introduce novel modifications to attention blocks, tailored to address challenges inherent to microscopy imaging, such as blurred boundaries and indistinct microorganism features. Our experimental results demonstrate significant performance improvements compared to baseline models, with the mean Intersection over Union (mIoU) increasing from 0.812 to 0.954 and the Dice coefficient improving from 0.864 to 0.968 on the Microscopy Diplococci Images Dataset (MDID). These advancements underscore the potential of specialized attention-based architectures to enhance diagnostic accuracy and operational efficiency in biomedical image analysis.