Cryo-FIB-SEM volumes stop drowning in charging artifacts: a two-stage restoration lets a standard FIB-SEM membrane predictor read them directly
Synopsis
Cryo focused-ion-beam scanning electron microscopy images cells in a near-native hydrated state in 3D, but charging artifacts, noise and weak membrane contrast have long hampered interpretation. A post-acquisition two-stage deep-learning pipeline translates cryo images toward the conventional room-temperature FIB-SEM domain: stage one suppresses large-scale artifacts slice by slice, and stage two uses five consecutive slices to enhance membrane contrast. On 164 held-out synthetic images, stage one raised mean paired PSNR by 16.74 dB over degraded inputs; across four experimental datasets, visible artifacts decreased and membranes became clearer, and the pretrained CebraNET membrane predictor could be applied directly to the processed volumes without retraining.
Interpretation
Cryo-FIB-SEM volumes can now be processed after acquisition into images with fewer artifacts and stronger membrane contrast, without changing the imaging hardware. Cryo specimens accumulate charge because they conduct poorly, producing dark cores, directional tails, stripes and uneven backgrounds, compounded by noise and loss of focus, which makes membranes hard to annotate; existing registration, denoising and artifact-correction methods must handle degradations that coexist across spatial scales and differ between adjacent slices. On 164 held-out 1024×1024-pixel synthetic test images, stage-one output showed a mean paired PSNR gain of 16.74 dB against constructed reference targets, with matching improvements in SSIM and MAE; 28 separate artifact-free identity controls showed the input was barely altered.
The two stages divide the work by spatial scale: first suppress broad artifacts, then enhance local membrane signal. Charging artifacts can extend laterally over large fields, while membranes carry meaningful correspondence across adjacent slices, so the two kinds of information need different spatial context. Stage one is a NAFNet-based 2D model that processes each full 1024×1024 slice independently; stage two is a five-channel U-Net that predicts an enhanced central 256×256 patch from five consecutive slices, with center-referenced local alignment of neighboring patches at inference.
Processed cryo volumes can be handed directly to a membrane segmentation model trained on conventional room-temperature FIB-SEM data. Analysis tools for cryo-FIB-SEM previously had to be developed separately, whereas conventional volume EM already has mature membrane segmentation and organelle detection pipelines. The pretrained CebraNET membrane predictor ran on original and processed volumes without retraining or fine-tuning and with the threshold fixed at 0.30; on the worm–algae symbiosis and mouse brain datasets, predictions on processed inputs followed visible cell boundaries more closely, including the stacked thylakoid membranes of algal chloroplasts.
Perspective
The demonstrated scope is held-out synthetic tests plus four acquired cryo-FIB-SEM datasets: the public Chlamydomonas reinhardtii dataset EMPIAR-11275 and three in-house datasets of human small airway epithelial cells, AppSAA knock-in mouse brain tissue and a worm–algae symbiotic system; all four were processed by the same two-stage inference procedure without dataset-specific retraining.
The synthetic tests measure agreement with constructed reference targets and are not pixel-level accuracy estimates on cryo-FIB-SEM data; when the central slice and its neighbors are severely obscured by charging or halos, the underlying morphology cannot be reliably inferred. Membrane segmentation gains are shown only qualitatively, because no independent manual membrane ground truth exists, so neither detection accuracy nor the correctness of reconstructed membrane topology is quantitatively established.
