Synthetic Dataset Creation to Train a Cryo-EM Segmentation Model via Critical Points

This research introduces an innovative synthetic dataset generation approach for cryo-electron microscopy (cryoEM) protein structure segmentation using automated critical point detection. Leveraging radial projection techniques and Fibonacci sphere sampling, we develop a synthetic data augmentation methodology that eliminates manual point selection while maintaining segmentation accuracy. When applied to the DeepEMSeg framework, our synthetic dataset approach demonstrated improved Intersection over Union (IoU) and Dice coefficient metrics across diverse protein structures. The results evidence that we can eliminate the need for manual point selection, offering an automated alternative to traditional human-guided dataset preparation approaches.