Improved Mitochondrial Semantic Segmentation with Fine-Tuned Multi-Modal U-Net

Mitochondria intricately regulate fundamental cellular processes, including energy production, signalling, differentiation, cell growth, and programmed cell death. Recent advancements in deep learning architectures have demonstrated superior performance in mitochondria segmentation using publicly available datasets. In this study, we propose a novel approach for refining mitochondria semantic segmentation, employing a Multi-Modal U-Net architecture. Our investigation systematically assesses the impact of three distinct loss functions, varied dataset sizes, and different training durations on segmentation performance. Leveraging the U-Net architecture's intrinsic capacity to capture fine-grained features and retain spatial information, we extend it with multi-scale processing. This augmentation enables the incorporation of both global contextual information and local details, resulting in enhanced accuracy and robustness in mitochondria segmentation. Evaluation metrics, such as the Mean Intersection over Union (mIoU) and mean Dice Coefficient Score (mDCS), are employed to gauge segmentation quality. Through extensive experimentation, we identify a finely-tuned configuration that consistently outperforms current state-of-the-art methods, highlighting notable advancements in both accuracy and efficiency.