Overview
Built a 5-stage weakly supervised deep learning pipeline combining cell-level CNNs and attention-based Multiple Instance Learning (MIL), processing 10K+ single-cell crops from 4-channel fluorescence images.
Problem
Cell classification from fluorescence microscopy images is expensive to label at the instance level. Weak supervision allows training with only slide-level labels, dramatically reducing annotation cost.
Approach
- Cell-level CNNs for feature extraction from single-cell crops
- Attention-based MIL for aggregating instance-level predictions
- Pseudo-label refinement guided by attention weights
- Per-class threshold tuning to handle class imbalance
- 4-channel fluorescence image processing pipeline
Results
| Metric | Score | |--------|-------| | Macro-F1 | 0.45 | | Weighted-F1 | 0.60 | | Single-cell crops | 10,000+ |
Tech Stack
Python PyTorch CNN Attention-MIL OpenCV NumPy