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

Deep learning label-free cell imaging has become essential in modern medical applications, enabling precise cell analysis while preserving natural biological functions and structures by removing the need for potentially disruptive staining reagents. However, developing accurate and generalizable models for these tasks requires a large amount of data. In this work, we propose LIVECell-CLS, the largest benchmark dataset for label-free cell classification to date, constructed leveraging the LIVECell segmentation dataset. LIVECell-CLS consists of over 1.6 million images across 8 distinct cell lines. We employ this dataset to perform a comprehensive comparison of 16 baseline deep learning models spanning diverse architectures like ResNets, ViTs, or MLP-Mixers. Our results indicate that models with locality inductive biases like CNNs and Swin-Transformers generally outperform patch-based models such as ViTs and MLP-Mixers in balanced accuracy and F1-score. We further propose Tensor Network variants of the baseline backbones, leveraging a C.elegans connectome-inspired module to improve the latent representation prior to the final classification. Such variants consistently improve classification performance across all architectures, with gains up to 4 percentage points in test accuracy and relatively low parameter overhead. Our best model, Elegans-EfficientNetV2-M achieves 90.35% test accuracy and 94.82% F1-score. Finally, through the application of multiple Explainable AI techniques and UMAP visualizations, we analyze the extracted cell features providing insights into how different models work on this type of images. Our results reveal that accuracy gains correspond to enhanced feature separability and precision in model decision-making, particularly in challenging scenarios involving morphologically similar cell lines. Code and data are respectively available at: https://github.com/NeuRoNeLab/connectome-livecell-cls and https://kaggle.com/datasets/3eb5e4d9b2d603944dfc1a85fd37c6ba61b2c08ee543bd37342544370a88c71d.