Robustness of architecture rankings under distribution shift

Characterize the performance and relative architecture ordering of EfficientNet and Vision Transformer models for label-free single-cell classification under cross-instrument, cross-laboratory, and cross-modality distribution shifts.

Background

The benchmark evaluates cropped phase-contrast microscopy images from the controlled LIVECell dataset, whose acquisition conditions and morphology are more standardized than those encountered in operational laboratory settings. The authors therefore caution that their conclusions may not generalize to data acquired with different microscopes, in different laboratories, or using different imaging modalities.

The unresolved issue is whether the apparent similarity or ordering of CNN and transformer performance persists under stronger domain shift, which is important for assessing whether the reported accuracy–efficiency conclusions support real-world deployment.