Search papers, labs, and topics across Lattice.
To resolve conflicting claims regarding CNNs versus Vision Transformers in label-free single-cell classification, the authors benchmarked EfficientNet, ViT, and EVA-02 architectures on LIVECell data under strictly controlled, source-image-disjoint splits. They show that prior CNN superiority claims are an artifact of pretraining disparities—which provide a 3–4 point macro-F1 boost—reducing the true performance delta between matched CNNs and transformers to under 0.5 points. Crucially, distilling ensemble teacher councils into a compact EfficientNet-B0 yields a model that outperforms every individually trained backbone, including its larger teachers.
The longstanding performance gap between CNNs and transformers in cell microscopy collapses to under 0.5 macro-F1 once pretraining is controlled, paving the way for distilled tiny students that outperform massive frontier backbones.
Choosing a deep learning architecture for label-free single-cell classification remains an open question, with microscopy benchmarks reporting conflicting conclusions about CNNs versus transformers. We present a controlled benchmark on LIVECell phase-contrast microscopy data using source-image-disjoint train/validation/test splits to prevent parent-image leakage and matched optimisation, augmentation, and evaluation protocols across EfficientNet, Vision Transformer (ViT), and EVA-02 models. This allows the effects of architecture, pretraining, fine-tuning, tokenisation, and distillation to be disentangled. We find that the previously reported CNN advantage is largely explained by pretraining rather than architecture: the smallest pretrained model outperforms the strongest model trained from scratch despite far fewer parameters. Pretraining improves macro-F1 by 3-4 points, while the gap between the best pretrained CNN and transformer is below 0.5 points. Architectural choices nevertheless matter: ViT-S/8 outperforms ViT-S/16 and matches the four-times-larger ViT-B/16 at a quarter of the parameters, showing that finer tokenisation benefits small cell crops. Conversely, layer-wise learning-rate decay, central to the EVA-02 fine-tuning recipe, degrades performance, highlighting that transfer heuristics from natural-image recognition may not generalise to microscopy. Finally, knowledge distillation substantially improves the deployment frontier: compact EfficientNet-B0 students distilled from teacher councils outperform every individually trained backbone, including the EfficientNet-B5 and EVA-02 teachers. Overall, our results show that rigorous control of pretraining and evaluation is essential for interpreting biomedical architecture benchmarks, while distillation may be a more effective route to practical single-cell classification than architecture choice alone.
