Deeptosis

Welcome to Deeptosis!

Deeptosis is an automated analysis pipeline for label-free identification of cell death phenotypes from bright-field microscopy images.

Users can upload raw images, and the system performs cell segmentation, single-cell extraction, and classification into apoptosis, pyroptosis, or other states, returning both annotated images and per-cell summary tables.

At the core of Deeptosis is a fine-tuned Vision Transformer (ViT) model that operates on single-cell crops. Each cropped cell image is divided into fixed-size patches and encoded through stacked self-attention blocks, enabling the model to capture subtle morphological differences associated with distinct cell death programs. The overall model design include patch embedding, transformer encoder blocks, and the classification head.

The classifier demonstrates high discriminative power across apoptosis, pyroptosis, and other states. In five-fold cross-validation and on an independent test set, the model achieves AUROC values above 0.98 for all classes, with minimal confusion between apoptosis and pyroptosis. Quantitative performance metrics include ROC curves and confusion matrix.