Evaluation of training data, ground truth and shape variability in the semantic segmentation of HeLa cells observed with electron microscopy

This paper describes the quantitative and qualitative evaluation of training data, ground truth and the variability of the shape of HeLa cells as observed with electron microscopy in the process of semantic segmentation. HeLa cells are a widely used immortal cell line, but the principles described here could apply to other cells. From a data set of 8, 000x8, 000x517 voxels, a region of interest (ROI) of 2, 000x2, 000x300 was cropped for quantitative evaluation with the qualitative evaluation performed on a larger section of interest. Patches of data and labels were used to train a U-Net architecture from scratch, which was compared also with a traditional image processing algorithm. The correctness of ground truth, that is, the inclusion of one or more nuclei within a region of interest was evaluated with a hand-drawn ground truth of the ROI. The impact of the extent of training data was evaluated by comparing results from 36,000 data and label patches extracted from the odd slices in the central region, to 135,000 patches obtained from every other slice in the set. Accuracy and Jaccard similarity index were higher for the larger training set with multiple nuclei in the ground truth. However, when the U-Nets were used to segment the section of interest beyond the ROI, results were not as good as expected, suggesting that the very complicated nature of the shapes of HeLa cells cannot be learned from a single cell. Data, code and ground truth have been made publicly released through Empair, GitHub and Zenodo.