Synthetic Data Creation of Neurons for Training Machine Learning Segmentation Algorithms Volume Electron Microscopy

Mapping synaptic connectivity at cellular resolution with volume electron microscopy (EM) enables mechanistic study of neural circuits, but it requires dense segmentation and extensive proofreading. Recent connectomics resources such as the MICrONS cubic-millimeter dataset co-register calcium imaging from ∼75,000 neurons with an EM reconstruction containing >200,000 cells and ∼523 million synapses [1]. Whole-brain connectomics has also reached the adult female Drosophila brain (139,255 neurons; ∼5×10^7 synapses) [2,3]. Despite progress in automated methods, generating sufficient high-quality labeled training data remains a major bottleneck. Volume EM datasets are commonly acquired by serial sectioning (e.g., automated tape-collecting ultramicrotomy, ATUM) followed by scanning electron microscopy (SEM) imaging and computational reconstruction [4]. While modern pipelines include automated segmentation, human-in-the-loop correction is often required to obtain reliable neuron instance segmentations at scale. Supervised machine learning models typically need large, diverse, and accurately labeled training sets; however, dense manual annotation for EM is expensive and difficult to scale. We therefore propose a synthetic-data pipeline that generates perfectly paired neuron instance masks and EM-like images to augment training for segmentation models, reducing dependence on manual labeling inspired by synthetic datasets used in PCB defect detection [5], Our approach is mask-first: (i) generate realistic neuron instance masks, then (ii) translate each mask into a matching EM-like image using conditional image-to-image translation. This pairing provides ground-truth labels by construction and is compatible with standard supervised training. As a foundational dataset, we downloaded 5,000 paired 512×512 patches from the publicly available MICrONS Explorer cortical mm^3 resource, including the corresponding neuron instance masks [2]. Patches containing fewer than two instances were excluded, yielding 3,800 paired samples for training and validation. For mask generation, we used Stable Diffusion (SD) and fine-tuned it via Low-Rank Adaptation (LoRA) on the 3,800 real instance-mask images. Captions specified “neuron instance mask” and the number of cells in each patch. LoRA training was performed for 10 epochs, and the best checkpoint was used to sample synthetic masks across a range of instance counts. For mask-to-SEM translation, we trained pix2pix and SPADE conditional generative adversarial networks (GANs) on real paired data (real mask → corresponding real SEM image). Synthetic SEM images were then generated by feeding synthetic masks into the trained generator. In our experiments, SPADE produced sharper boundaries and more plausible ultrastructural texture than pix2pix (Fig. 2) and was selected for the final pipeline. To prepare training data for Micro-SAM, each mask was converted to a single-channel integer label image (background = 0; each neuron assigned a unique ID) and saved as a 16-bit TIFF, paired with the corresponding SEM image saved as TIFF. We compared Micro-SAM models trained on (a) real-only data, (b) mixed real+synthetic data, and (c) synthetic-only data, evaluating on a held-out set of real MICrONS patches using qualitative overlays (Figs. 3–4) and standard segmentation metrics (Dice/IoU and instance-level precision/recall). Preliminary results indicate that synthetic-only training is insufficient, while synthetic augmentation provides additional diversity; remaining failure cases are associated with blurred boundaries and missing nuclear/organellar cues in the generated SEM images. Overall, these results suggest that realistic instance mask generation is achievable, while mask-to-image translation is currently the dominant limitation. Future work will focus on improving translation fidelity using stronger conditioning and domain-specific constraints (e.g., membrane continuity and texture statistics), and on a controlled low-label study to quantify whether synthetic augmentation reduces the amount of real labeled data required for robust performance. Example synthetic SEM-like image generated by translating a diffusion-generated neuron mask using a conditional GAN. Representative real SEM image patch from the MICrONS dataset used as target data for mask-to-image translation [1]. Example of good neuron instance segmentation on a real MICrONS SEM patch after Micro-SAM fine-tuning: original image (left), predicted instances (middle), and reference instances (right). Example failure case showing poor agreement between predicted and reference instances on a real MICrONS SEM patch (same layout as Fig. 3).