ARCTiC

Cryo-electron tomography reconstructs macromolecular structures inside cells by imaging a specimen at many tilt angles. In practice a sizable fraction of those individual tilt images is unusable: specimen drift blurs them, ice contamination and grid bars block the beam, sub-vitreous ice produces reflections that break alignment, and at high tilt angles the effective thickness of a milled lamella washes out contrast. Leaving these tilts in place degrades tilt-series alignment and contrast transfer function estimation, which propagates into every downstream step. Removing them is standard practice — and, in tools such as the RELION-5 exclusion job, entirely manual, costing minutes of inspection per tilt series and varying with whoever is doing the inspecting.

ARCTiC (Automated Removal of Corrupted Tilts in Cryo-ET), from Tomáš Majtner and Beata Turoňová at the Max Planck Institute of Biophysics, replaces that judgement call with a trained image classifier. The study benchmarked 13 ImageNet-pretrained architectures — CNNs and vision transformers — fine-tuned on 435 manually annotated tilt series from six in situ lamella datasets, and released the best four checkpoints for direct use. Alongside cataloged cryo-ET tools that address later stages of the pipeline, such as CryoLens for subtomogram representation, ProPicker for particle picking, and MemBrain v2 for membrane segmentation, ARCTiC targets the quality-control step that precedes all of them.

#Key Features

• Two classification modes: A binary model separates good from corrupted tilts; a six-class model additionally names the artifact as drift, ice reflection, lamella edge, thick lamellae, or contamination, so a user can discard only the categories that matter for their reconstruction.
• Ready-to-run checkpoints: Fine-tuned Swin_tiny, Swin_large, ResNet-50, and EfficientNet_b3 weights are published for both modes and applied to new tilt series without retraining.
• Fast enough to sit in the pipeline: Swin_tiny evaluates a 50-tilt series in 2.0 seconds on a GPU and 3.4 seconds on CPU; the largest model takes 5.1 seconds on GPU.
• Cheap adaptation to a new lab: Adding just four annotated tilt series to the fine-tuning set raised F1 for every tested architecture, and fine-tuning on four series takes under thirteen minutes on a single GPU.
• Interpretable predictions: Swin attention maps localize the image region driving a call, concentrating on lamella edges, contamination, and ice reflections.
• Reviewable output: A command-line script emits the cleaned tilt stack plus a PDF of the removed tilts and their predicted corruption probabilities.

#Technical Details

All networks start from ImageNet weights with the classification head replaced, then fine-tune with Adam at a learning rate of 0.0001, batch size 32, cross-entropy loss, and 50 epochs at 224 × 224 input resolution (320 × 320 for the EfficientNet variants). The corpus contains 2,621 corrupted and 22,726 good tilts across 435 series, split by series into 315 training, 29 validation, and 91 test series; horizontal flips and 90° rotations expand the corrupted training set eightfold to counter the class imbalance. On the results tables, Swin transformers lead the F1 ranking, while CNNs lead on AUC: EfficientNet_b3 records the highest binary AUC at 0.9049, ahead of ResNet-50 at 0.8945 and Swin_large at 0.8636, and ResNet-50 leads the multiclass task at 0.9673 with the Swin models at 0.9590–0.9602. Scored as a binary decision, the multiclass Swin_large model agrees with manual annotation on 93.4% of good tilts and 70.3% of corrupted ones; ResNet-50 catches more corrupted tilts (82.6%) at the cost of discarding more good ones. Drift is the hardest class for every model, and the transformers in particular recover it poorly. Trainable parameter counts range from 12.2M for EfficientNet_b3 to 195M for Swin_large.

#Applications

ARCTiC drops into an existing cryo-ET preprocessing pipeline immediately before tilt-series alignment. It is aimed at structural biology groups running high-throughput in situ tomography, where manual tilt curation of hundreds of series is a genuine bottleneck, and at multi-user facilities that want a single objective exclusion criterion instead of per-user habits. Because the released checkpoints fine-tune cheaply, groups working on in vitro specimens or non-milled cells can adapt them to data that differ from the in situ lamellae used for training.

#Impact

The practical contribution is standardization: on the test series, alignment residual means from automatically cleaned data closely track those from manually cleaned data, so the automated decision is a workable substitute rather than a compromise. The work was published in Journal of Structural Biology: X and the code is released under GPL-3.0. Scope is deliberately narrow — this is a per-tilt quality classifier, not a reconstruction or picking model — and the training labels carry the subjective biases of the six annotators who produced them, which the authors identify as the main target for future semi-supervised work.