Advancing label-free cell classification with connectome-inspired explainable models and a novel LIVECell-CLS dataset

Area of Science:

• Biomedical Imaging
• Computational Biology
• Artificial Intelligence

Background:

• Label-free cell imaging is crucial for medical applications, avoiding disruptive staining.
• Accurate cell analysis models require extensive, high-quality datasets.
• Existing datasets may not be sufficient for training robust label-free cell classification models.

Purpose of the Study:

• Introduce LIVECell-CLS, the largest benchmark dataset for label-free cell classification.
• Compare the performance of diverse deep learning architectures on this dataset.
• Investigate the efficacy of Tensor Network variants and explainable AI for cell classification.

Main Methods:

• Constructed LIVECell-CLS with over 1.6 million images across 8 cell lines from the LIVECell segmentation dataset.
• Evaluated 16 baseline deep learning models (CNNs, ViTs, MLP-Mixers) and proposed Tensor Network variants.
• Applied Explainable AI techniques and UMAP visualizations for feature analysis.

Main Results:

• Models with locality inductive biases (CNNs, Swin-Transformers) outperformed patch-based models (ViTs, MLP-Mixers).
• Tensor Network variants consistently improved classification performance across architectures, achieving up to 4% accuracy gain.
• The best model, Elegans-EfficientNetV2-M, reached 90.35% test accuracy and 94.82% F1-score.

Conclusions:

• LIVECell-CLS is a valuable resource for advancing label-free cell classification research.
• Locality-biased models and Tensor Network enhancements offer improved performance.
• Explainable AI analysis confirms accuracy gains correlate with enhanced feature separability.